The US Senate fell 11 votes short on September 15 of advancing the CLARITY Act, the American crypto market-structure bill, and CIOs and CFOs with a 2027 blockchain budget now face the same question: which rule, exactly, is the spending waiting on? For three of the four use cases mapped below, an existing US law or a US agency decision sets the date, and no new bill is needed. Only token markets still wait on the US Congress.

What the Senate Vote Stopped

Senate roll call vote 234 asked whether the US Senate would end debate on the motion to proceed to H.R. 3633, the CLARITY Act. It failed 49 to 50, with 60 of 100 votes needed. No Democrat voted yes, and Republican Senators Collins, Hawley, Moran and Tillis voted no. Tillis voted no to preserve a motion to reconsider, which he then filed. The law firm Orrick attributes the failure to disputes over ethics rules for officials’ crypto holdings and over stablecoin yield.

Thirty-eight days on, no later floor action has surfaced in the sources reviewed. Davis Wright Tremaine wrote on September 21 that “we do not foresee a new bill being able to garner 60 Senate votes in the near- or mid-term.” The 119th Congress ends January 3, 2027, and a bill that has not passed by then starts over.

A Gating Map for the 2027 Budget

Token markets are the only row where the stalled bill sits on the critical path, and the SEC’s proposed Regulation Crypto Assets does not change that. It covers offerings of crypto assets sold as investment contracts, not trading platforms, brokers or custody.

Stablecoin Payments: A Fixed Date, Then an Approval Clock

On October 21 the OCC finalized its GENIUS Act rule, and its release states that the rule and the Act take effect January 18, 2027. That date can no longer move earlier. The statute starts on the earlier of 18 months after enactment or 120 days after the primary regulators issue final rules, and 120 days before January 18 was September 20. The final rule arrived 31 days after that cutoff.

The OCC is not accepting applications yet; it says it will start once the Office of Management and Budget approves the paperwork. Under the final rule, the OCC has 120 days from receiving a substantially complete application to decide, and an application it has not denied by day 120 counts as approved. A substantially complete application received on January 18 would therefore get its answer by May 18, 2027.

Issuers with a substantially complete application pending can seek a waiver of up to 12 months from January 18. OCC Bulletin 2026-51 says the agency generally expects to grant them for issuers already operating in the US but sees no compelling case for those that are not.

The Federal Reserve was the last primary regulator to propose rules, and comments on its two proposals and on Treasury’s state-certification rule close November 30, Sullivan & Cromwell reports. Platforms that sell stablecoins to US customers face a separate cutoff on July 18, 2028, after which they generally may offer only stablecoins from permitted issuers.

My take: the gate a treasurer should plan around is issuer approval, not the Senate. For OCC-supervised issuers, no application can be filed yet.

Yield: Where the Failed Bill Still Bites

GENIUS bars issuers from paying interest or yield solely in connection with holding, using or retaining a payment stablecoin. The open question is whether exchanges and other third parties can pass reserve income to users. The OCC’s proposal included a rebuttable presumption that some issuer arrangements with affiliates and related third parties broke the ban; the final rule drops it, and the OCC says it retains its authority to address evasion. A regulatory official said such cases will be handled one at a time, The Block reported.

The American Bankers Association accused the OCC of undermining Congress. The Bank Policy Institute’s Paige Pidano Paridon said the omission “replaces clarity with ambiguity,” while the Blockchain Association said dropping it “brings the final rule more closely in line with the statutory text.”

Yield was also one of the disputes behind the CLARITY stall. The implication for a 2027 business case is narrow but real: payments and settlement do not need yield, but treasury products that pay a return on idle balances do, and a model that books that return today assumes an answer nobody has given.

Banks and Securities Venues Did Not Wait

Tokenized deposits already sit under existing law. The FDIC’s April proposal, in the Federal Register, would confirm that “tokenized forms of deposits are not a separate category of deposits under the statute.” JPMorgan rolled out its JPMD deposit token on Base for institutional clients in November 2025, The Block reported. On July 9, Swift said its ledger was “ready for initial use.” By September 30, 83 days later, ANZ had run live US dollar payments for BHP between Melbourne and New York with Citi.

The limits show too. Citi’s September 2 release calls its live pilot transactions on the ledger part of a proof-of-concept phase running from July to December, with final settlement on existing models such as real-time gross settlement. ANZ did not disclose payment values, and BHP’s Marlon Singh called the work “still at an early stage.” Wells Fargo said on August 4 that tokenized deposits would roll out “this fall” for select corporate clients; no launch announcement had surfaced by October 23.

Tokenized stocks took a separate road. The SEC’s September 17 order gives five-year conditional relief to venues that trade tokenized versions of listed stocks without registering as exchanges. Decrypt reported that an OKX and ICE joint venture published a notice dated October 4 for a 24/7 venue covering more than 60 US stocks, with no launch date, and The Block counted nine objecting issuers by October 17, including Broadcom, CoreWeave and Verizon. SIFMA’s October 19 letter asks the SEC to promptly address issuer notice-and-consent and beneficial-ownership questions, among four areas, and says it would support more time before venues open.

Testing the Three Assumptions

The first assumption, that adoption hinges on clarity, survives in part. PYMNTS Intelligence’s 2026 Certainty Project surveyed 60 US middle-market CFOs: 67% cited regulatory or compliance uncertainty as a barrier to stablecoins. EY-Parthenon’s survey of 350 financial-institution and corporate respondents, fielded in June 2025 before GENIUS became law, found 73% naming regulatory uncertainty among their top three concerns. Yet only 13% of the PYMNTS CFOs use stablecoins today, and 45% said bank integration would make stablecoins a more meaningful part of payment flows, while 40% cited regulatory clarity. Neither survey measures the CLARITY vote, and the PYMNTS sample is small.

The second, that institutional investor warmth signals enterprise readiness, finds no support in the sources reviewed. Taurus co-founder Lamine Brahimi told CoinDesk that a bank joining Swift’s ledger needs its own permissioned ledger, wallet capabilities, and tokenization and smart-contract tooling. My read: those are capabilities a bank must build or buy, and none of them requires a regulatory approval. Taurus sells custody and tokenization infrastructure, so weigh the source.

The third, that year-end budgets key off regulation, has no evidence behind it. No survey or named company reviewed for this piece ties 2027 blockchain budgets to the CLARITY vote. The CFO surveys measure barriers; they do not date spending decisions.

What Could Still Go Wrong

Agency action moves faster than legislation and is easier to undo. Orrick notes that agency rules are more open to judicial challenge and reversal, and the SEC’s tokenized-stock order is temporary by design. The yield dispute shows the stablecoin rulebook is not settled either.

The CFTC sent its crypto market rules to White House review on September 17, then issued an advance notice of proposed rulemaking on October 5, with comments due December 14. CoinDesk notes the effort may leave “a significant gap” because the CFTC lacks authority over spot markets, and Davis Wright Tremaine expected in September that binding rules would come no sooner than late 2027. For that row of the map, the claim that adoption hinges on regulatory clarity holds.

The Signal to Watch

The first signal worth watching is the OCC announcing that it is accepting applications, because the 120-day decision clock cannot start before then. January 18 falls 15 days after this Congress ends, so the stablecoin law takes effect whether or not the market-structure bill returns. The next Senate vote will matter to token markets, but it does not control the date the stablecoin law starts.

The post CLARITY Stalled in the US Senate: Enterprise Stablecoin Payments Still Have a Start Date appeared first on DataFLOQ.

What php2shell Is

php2shell is the name for two vulnerabilities, CVE-2026-109618 and CVE-2026-109619, each rated 9.2 on the CVSS v4.0 scale by PHP’s security team. In the first, PHP runs out of memory partway through a request and leaves some internal data uninitialized, and request cleanup may then treat that data as valid. In the second, internal state from one request can outlive the memory behind it, and a later request handled by the same worker may use that stale state before PHP resets it. Cloudflare says the flaws may allow an attacker to cause denial of service, remote code execution and data exposure.

PHP says all installations serving web requests are at risk, including PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP and LiteSpeed, and it considers the risk highest for mod_php. The advisory for the first flaw adds that no complete exploit has been demonstrated against PHP-FPM, and its CVSS vector rates attack complexity as high. The PHP Foundation said its security team did not identify any functional exploits at the time of release, although proof-of-concept code had begun to circulate.

Cloudflare says Calif.io, Wiz, OpenAI and Quarkslab disclosed the flaws. Wiz says it reported its findings to PHP on August 24, that PHP acknowledged them on September 21, and that AI models helped throughout the research. PHP shipped fixes on October 21 in versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Cloudflare says the flaws affect all web servers running PHP 5.4 or later. W3Techs puts PHP at 69.3 percent of the websites whose server-side language it can identify, as of October 22, and Cloudflare names WordPress and Magento among the platforms that depend on it.

Why a WAF Has to Read the Response

An application cannot defend itself against these flaws. Datadog’s reproduction used an unauthenticated POST request that never reached application code, so application-level authentication does not block it. A firewall in front of the server is one of the few places to intervene before patching, but request inspection has limits. PHP’s advisories list multipart techniques among the known ways to trigger the out-of-memory condition and name session upload progress, which is enabled by default, as a trigger for the second flaw. Cloudflare says request data alone does not always provide enough evidence to distinguish an exploit from legitimate traffic.

Cloudflare’s answer is full transaction detection, which it described in a companion post the same day. The capability evaluates a request together with the origin’s response, and when a blocking rule matches, Cloudflare replaces the origin response with a WAF block page or a custom response. The php2shell protections are the first managed detections Cloudflare has released with it. The companion post’s example is a path traversal request for a database configuration file: request inspection can recognize the attempt, but the request alone cannot establish whether the application would expose the file. The response is where that answer shows up.

The design has a limit that Cloudflare states itself. A response rule cannot undo actions already performed at the origin. The implication is that response inspection protects data leaving the server better than it protects the server. Imperva describes one php2shell technique that returns about 67 KB of server memory in a Set-Cookie header carrying the session ID, the kind of leak response-side detection targets. Code that has already run on the origin is another matter. Cloudflare has not published what its php2shell signatures look for, so any link between them and that leak is inference.

What Each Plan Gets

Cloudflare says free and paid WAF customers were protected by default from 03:02 UTC on October 21. Free customers need to do nothing, because the rules sit in the Cloudflare Managed Free Ruleset. Professional, Business and Enterprise customers get seven rules in the Cloudflare Managed Ruleset, and Cloudflare asks them to review the deployment and confirm it runs in block mode. Every plan, free included, also gets a response-side rule in the Cloudflare Managed Response Firewall Default ruleset that blocks high-confidence signatures by default.

The labels are generic. The rules read “Anomalous Execution” followed by a number from one to seven, with no mention of PHP or the CVEs. Cloudflare’s post says the rule descriptions will be updated soon and lists the rule IDs so teams can filter matches under Security Analytics. Until the labels change, a security team scanning its dashboard cannot tell from the names which rules cover php2shell.

The Clock Started Before Disclosure

The sequence matters. CVE identifiers began circulating online around 15:00 UTC on October 20 with a wrong release time and a wrong CVSS score of 10.0, according to the PHP Foundation, which says no technical details had leaked. PHP posted an advance notice at 19:33 UTC and moved its release up from October 22 to 14:00 UTC on October 21. The foundation adds that Calif then accidentally published a video of its proof of concept, which helped others develop their own proof-of-concept code faster.

Cloudflare’s rules went live at 03:02 UTC on October 21. Cloudflare counts its lead as about nine hours, but PHP’s release landed at 14:00 UTC, which makes the gap closer to eleven. Cloudflare says it had been developing protections since early October, working with the PHP Foundation, OpenAI, Wiz and others. It was not the only vendor with early notice. PHP says Linux distributions and LTS vendors received the patches in advance, and Imperva says the researchers briefed it confidentially before disclosure and that its Cloud WAF customers are protected out of the box.

Attackers did not wait for the leak. Amazon Threat Intelligence reported that its MadPot honeypot logged an exploitation attempt dated October 17, four days before disclosure, and said it remains unclear how the actor obtained exploit information. Cloudflare says it observed attack traffic against its customers and saw attackers mutating payloads to get around its rules, though it published no counts.

Vendors describe attempts more than successes. watchTowr said on October 21 that it was observing active exploitation attempts against both flaws and did not say any succeeded. Amazon says php2shell is now being broadly exploited in the wild, while also saying it is unclear whether the attempts it saw would succeed on real systems. The Hacker News reported that no source had confirmed a successful break-in of a specific server as of its October 21 update. For an unpatched server outside Cloudflare’s protection, the earliest known attack attempt predates disclosure by four days, and Amazon’s guidance reflects that: start investigations from the earliest known exploitation, not from the disclosure date.

Patching Is Still the Only Complete Fix

Cloudflare says plainly that customers should patch PHP even with the WAF in place, and it describes its rules as an additional layer of defense in depth. The flaws sit in the PHP runtime, so updating website code or plugins does not fix them, and customers on managed hosting should ask their provider to apply the security update. PHP’s release notes say there are no complete mitigations and that updating is the only safe choice. A WAF also inspects only the traffic routed through it, so an origin server reachable by another route gets no cover.

Patching carries its own friction. PHP 8.1 and older no longer receive updates from the project, and 8.2 reaches end of life on December 31, 2026. For 7.4, 8.0 and 8.1, the security team provided minimal source backports for people building from source or without an LTS vendor.

Laravel Forge warned customers that its automatic security updates do not patch PHP, because Forge installs PHP from a dedicated package repository outside Ubuntu’s own security channel. Long-running processes such as queue workers keep the old PHP until someone restarts them.

What Operators Should Take From It

The signal worth watching is the response side. Cloudflare says it will expand coverage over time to include detecting data exposure and exfiltration signals, which suggests its managed rules are moving from judging requests toward judging what servers send back. php2shell fits that pattern, because on the memory-disclosure path the evidence shows up in the reply.

The implication for operators is straightforward. Inventory every PHP version in production, patch the supported branches, plan a move off 8.1 and earlier, and restart long-running workers after the update. Amazon also advises searching form-upload POST requests for the string PHP_SESSION_UPLOAD_PROGRESS and checking servers for unexpected process execution or file changes. The WAF lowers the odds while that work happens, and Cloudflare itself says attackers are already reshaping their payloads against it.

Any similar runtime flaw will test the same arrangement: vendors with advance notice, a leak that shrinks the window, and customers who still have to do the patching. Clearer rule names and published attack counts would let those customers judge how much time a head start bought them.

The post Cloudflare Deployed php2shell Protection Hours Before Disclosure. Here’s What the Head Start Doesn’t Cover appeared first on DataFLOQ.

A Power Event in One Zone, Errors Across the Region

Google’s status page logged the first problems at 06:08 UTC on 20 October, which was 07:08 in London. The initial notice described “elevated errors and intermittent network failures” in the europe-west2 region. Google later named the cause: “A power event caused an outage within europe-west2-b zone,” one of three zones in its London region. Affected services included Google Kubernetes Engine, Cloud Run, App Engine and Cloud Filestore.

Recovery came in stages. Google said the power event was mitigated and that it was prioritizing existing workloads, while warning that customers might not yet be able to create new workloads. It told customers to migrate affected workloads to another zone or region “if feasible.” By 14:03 UTC, according to The Register, Google said the issue was mitigated for most customers and that a very small number remained affected. It promised an analysis once its internal investigation ends, and had given no detail on what caused the power event itself.

Revolut Confirmed the Link, the Banks Have Not

Revolut tied its disruption to Google in its own words: “Due to a recent Google Cloud outage, customers may have experienced temporary service disruption on the Revolut app today.” Its status page flagged card payments, transfers, ATM withdrawals, currency exchange and stock trading. Euronews reported users hit in the UK, Ireland, France, Spain and Japan. Revolut also said customers’ money and data were safe.

Lloyds, Halifax, Bank of Scotland, HSBC and RBS all appeared in outage reports, but the evidence for a Google link is thinner. A Lloyds spokesperson told The Wall Street Journal that the bank was aware of similar issues affecting other organizations, and the Journal reported that its app stayed accessible while some web services had problems. Bloomberg reported problems at HSBC and RBS, and neither bank responded to its inquiries.

The courts offer a third case. The Criminal Bar Association said the Cloud Video Platform, which courts use for remote hearings, had been down across England and Wales since 10 a.m. UK time, and that many barristers could not proceed with trials and hearings. In Llanelli, court records gave “CVP DOWN NATIONALLY” as the reason two hearings were adjourned, Herald.Wales reported. The Ministry of Justice said by late afternoon that court systems were still not fully restored, according to Computer Weekly. No statement reported so far shows HM Courts and Tribunals Service tying the failure to Google Cloud, so that link rests on timing.

Mitigated Is the Vendor’s Word

Revolut’s recovery ran well past Google’s mitigation. Reuters reported that at 17:00 GMT only six of Revolut’s 27 listed services were working, support chat and internal transfers among them, while card payments, ATM withdrawals, currency exchange and stock trading were still disrupted. That was nearly three hours after Google’s 14:03 UTC update and almost 11 hours after the first error. Revolut said that evening that the issue was fixed and services were restored. One reading, consistent with Google’s own warning about creating new workloads, is that a customer’s recovery depends on rebuild work that starts only after the vendor stabilizes the zone. Nothing in the reporting so far confirms that, so it remains an inference.

Google’s account has a defensible core: the root cause sat in one of three zones, and by Google’s telling most customers were back by about 3 p.m. London time. Its wording did shift, though. The first notice described errors across the region, while the confirmed cause was one zone.

Zones exist to contain failures like this one, so the open question is why a single-zone power event reached so many services and customers. London has raised that question before. In July 2022, a heatwave cooling failure caused an outage of 18 hours and 23 minutes in europe-west2-a, and Google’s incident report said it had inadvertently changed traffic routing for internal services to avoid all three zones in the region, The Register reported. Whether Tuesday followed the same pattern is not yet known.

London Is Google’s Only UK Region

Google’s published location lists show London (europe-west2) as its only UK region. A firm can spread UK workloads across the three London zones, but it cannot spread them across UK regions. Google’s advice to move workloads elsewhere “if feasible” therefore points those firms toward a region in another country. Whether that is permitted depends on each firm’s data residency and regulatory obligations, which a status page cannot settle.

The Oversight Regime Is Three Months Old

On 13 July, HM Treasury’s designation of four providers took effect: Google Cloud EMEA Limited, Microsoft Ireland Operations Limited, Amazon Web Services EMEA SARL and Oracle Corporation UK Limited. Under the UK’s critical third parties regime, the Prudential Regulation Authority, the Financial Conduct Authority and the Bank of England can make rules for the services those providers supply to regulated firms, demand information and, in law firm Sullivan & Cromwell’s summary, “direct a CTP to take certain actions or to cease an activity.” The EU moved first: in November 2025, the European Supervisory Authorities named 19 providers, Google Cloud’s European arms among them, as critical under the Digital Operational Resilience Act, FStech reported.

The UK timetable is gradual. A designated provider owes an initial self-assessment within three months, a mark that arrived roughly a week before the outage. Mapping, first-round scenario testing and an incident management playbook exercise are due within 12 months. Designation also “does not remove the responsibility of regulated firms” for their own third-party risk, Sullivan & Cromwell notes.

Scrutiny arrived the same day. The Treasury Committee published a letter on 20 October from its chair, Dame Meg Hillier, to the Governor of the Bank of England and the chief executive of the FCA, asking how they will use their powers in response to the incident. An FCA spokesman told The Telegraph that the regulator was “aware of an incident affecting services provided by Google” and was engaging with Google and affected firms. The Bank of England and the PRA declined to comment. The date also echoes last year: twelve months earlier to the day, a 14-hour AWS outage in US-East-1 disrupted UK services including Lloyds and HMRC, Computer Weekly reported.

Vendor Status Is Not Customer Recovery

The implication is that a hyperscaler’s status page measures the hyperscaler’s recovery. A bank or a payments app has to measure the moment its own customers can transact again, and Tuesday’s timeline suggests the two can sit hours apart. Firms that have not rehearsed losing a zone, and timed the recovery from the customer’s side, will learn that number during an incident instead.

Two documents should settle most of what remains open. Google’s promised analysis should explain why a single-zone power event produced errors across the region, and the regulators’ reply to the Treasury Committee should show what the new powers allow in practice. Until then, Revolut’s timeline shows how long “mitigated” can take to reach a customer.

The post Google Cloud Calls Its London Outage Mitigated. Revolut Customers Waited Hours Longer appeared first on DataFLOQ.

Where the Main I/O Enterprise Items Stand

I/O was not where Google launched its enterprise agent platform. That came on April 22 at Cloud Next, when Google introduced the Gemini Enterprise Agent Platform. Google called it a new developer platform and an evolution of Vertex AI that brings together models, development and tuning services. Model customization followed: supervised and reinforcement learning fine-tuning for Gemini 3.5 Flash both entered public preview in June. The Gemini Enterprise app, where employees find and run agents, is built on that platform.

Google Cloud CEO Thomas Kurian’s I/O post added a second wave. The table shows where seven of its items stood on October 16, based on Google’s release notes and announcement posts.

The Control Layer Shipped First

Agent Gateway and Agent Registry both reached general availability on June 18. Gateway governs connections between users, agents and tools. Model Armor guardrails for it followed on June 24. Controls kept arriving through the summer. The Antigravity developer tools that launched on August 18 let administrators set file-access and terminal-command policies. On September 8, context-aware access for Gemini Enterprise let them restrict use by device security and location.

Kurian’s I/O post offers one reason that order would make sense. Each Spark task runs in a fresh, isolated virtual machine, he wrote, and all traffic routes through Agent Gateway, which enforces data loss prevention policies. The implication is that Spark depends on Agent Gateway, which reached general availability in June. Google has not said it held Spark back for that reason, so the sequencing is an inference.

I/O’s Workspace Features Arrived Late and Narrow

In May, Google said Pics, its AI image tool, would reach business customers in preview “this summer.” It became generally available on September 1 for Business Standard and Plus and Enterprise Standard and Plus, starting with Rapid Release domains. Pics is on by default, and administrators can turn it off by domain, organizational unit or group. Google also says users get higher access to generative AI features in Pics at least through February 28, 2027, and that access may be limited afterward.

Voice took longer. Google launched Docs Live, Keep Live and Gmail voice on September 3 for consumer Google AI subscribers and said all three were “coming soon” for business customers. Gmail Live began rolling out to paid Workspace plans on October 15, with limits: mobile only, and only for U.S. users whose language is set to English (US). It is on by default for organizations with Gemini in Workspace access. Google’s Workspace Updates blog shows no business rollout for Docs Live or Keep Live through October 16.

Voice and Spark both reached consumer subscribers first. Google’s “this summer” turned into September 1 for Pics and October 15 for Gmail voice, and “soon” for Spark still has no date in the release notes.

Four Newer Gemini Models Since May

Gemini 3.5 Flash became generally available in the Gemini Enterprise app on May 19. The release notes have since added Gemini 3.6 Flash on July 21, 3.7 Flash on August 13, 3.8 Flash on September 2 and Gemini 4 Argon on October 13. Google turned 3.8 Flash on by default in the app on September 23. On July 21 it also announced that 3.5 Flash would leave the app’s global region on August 4. It postponed that on August 6 and said a revised removal schedule “will be provided when available.”

The summer upgrades came with administrative work. Gemini 3.6 Flash reached the global region on July 21, and administrators had to switch it on. In the U.S. multi-region, which offers data residency at rest and in processing, it stayed limited to allowlisted projects from July 24 until general availability on August 18. Administrators in unsupported regions could enable it by accepting a warning that traffic would route to the global endpoint, which does not support regional data residency.

Developers had changes to absorb as well. According to the Agent Platform release notes, the Gemini 3.6 Flash API ignores custom temperature, top-K and top-P values, returns an error for custom penalty settings, and rejects requests that end on a model turn.

Later releases handled residency better. Gemini 3.7 Flash, 3.8 Flash and Gemini 4 Argon each reached the global, U.S. and EU regions at once.

My take is that the model schedule, more than the product roadmap, belongs in the contract conversation. Buyers should ask for a written model-retirement policy, a residency timeline for each new model, and a compatibility commitment for API changes.

Gemini 4 Argon is off by default in the Gemini Enterprise app, and its toggle switches to on from October 27. For Spark, the release notes offer nothing to put on a calendar yet.

The post Google Said Its Enterprise AI Agent Was Coming “Soon”, Five Months On, the Guardrails Shipped First appeared first on DataFLOQ.

What IBM Announced

IBM published its announcement on October 1, three days after Gartner released the report on September 28. Analysts Shameen Pillai, Nicholas Carter and John Santoro assessed 17 vendors, the same number as in the 2025 edition. IBM’s post centers on IBM API Connect and its DataPower gateways, and frames the result around “the age of AI.”

IBM describes API Connect as combining AI-assisted development, full lifecycle API management and runtime governance across hybrid environments. A feature called Flow Pilot lets developers use AI coding agents from their preferred coding assistants to design, build, test and publish APIs. DataPower Interact Gateway, in IBM’s words, turns existing APIs and integrations into “AI-ready tools that agents can securely discover and invoke,” while DataPower Nano Gateway enforces policy from a lightweight, Kubernetes-native component placed close to workloads. IBM also describes federated management that applies consistent governance across IBM and third-party gateways.

The post does not say which of these capabilities are generally available. Its only forward-looking line is that IBM “will continue to invest across the API lifecycle.” Gartner’s disclaimer, reproduced in the post, says it does not endorse any vendor depicted in its publications.

Six Leaders Announced the Same Result

Every placement below comes from the vendors’ own announcements. Gartner’s public abstract lists no vendors, and the full report is gated behind Gartner access or vendor registration forms.

The AI messaging is nearly uniform. Kong says its AI Gateway 2.0 release adds governed access for agents and MCP tools. Boomi CEO Steve Lucas said enterprises “are no longer treating the API layer and the agent layer as separate problems.” MuleSoft credits unified API and AI management through its Omni Gateway, and WSO2 describes its placement as validation of its approach to AI governance. Each is a vendor describing its own product.

Gartner’s Wording Gives the Pitch Some Cover

The vendors are not inventing the premise. The 2026 Magic Quadrant abstract opens with “APIs provide the foundation for AI initiatives, headless architectures, modernization and digital product monetization,” then adds that “their widespread use introduces significant challenges.” The companion Critical Capabilities abstract says API management platforms “now enable both conventional API use cases and advanced agentic and generative AI scenarios.” Its contents list on Gartner’s site includes entries such as AI Gateway Support, MCP Development and Management, and AI and Agent Management.

Gartner forecast in March 2024 that by 2026, more than 30% of the increase in demand for APIs would come from AI and tools using large language models. The figure is a forecast, not a measurement of current agent traffic.

IBM’s Use-Case Results Are More Mixed

Gartner’s Critical Capabilities report scores platforms against five use cases instead of placing them in a quadrant. IBM’s product pages say it ranked first in two of the five use cases in 2026 and among the top five vendors in all five. The same pages say it ranked first in five of six use cases in 2025. The number of use cases fell from six to five between editions, so the counts do not compare cleanly, and IBM’s public materials do not name the two use cases it won this year.

A rival has staked a claim in the use case closest to IBM’s pitch. Gravitee says it tied for the highest score in AI and Agent Management, though its page does not show the underlying scores. By IBM’s own count it sits in the top five for that use case too, so the question is the size of the gap, and the public material does not answer it.

What Buyers Should Ask Before Trusting the Badge

My take is that an 11-year streak reflects sustained scores on Gartner’s two axes, Ability to Execute and Completeness of Vision, and a single quadrant position cannot show how a given Leader performs when an agent, not a developer, is the one calling the API. Buyers should ask every Leader for its use-case scores and for the release status of its MCP and agent-governance features before treating the label as an answer.

For organizations that already run gateways from more than one vendor, IBM’s federated management claim is the part of its pitch to test first, because it targets the estate they have instead of asking them to replace it.

Gartner’s forecast date for AI-driven API demand is this year, and the six vendors above have all said they are ready for it. Buyers who press for release status and use-case scores now will not need to wait for the 2027 edition to learn which of those claims hold.

The post IBM Is a Gartner API Management Leader for the 11th Time. Now the Test Is AI Agent Governance appeared first on DataFLOQ.

What Happened, and How Big It Is

Retailers confirmed the attack first. Dominique Schelcher, chief executive of Coopérative U, posted on Monday that cyberattacks had hit the refrigeration equipment of several stores, and he called the attack “unprecedented” and “unacceptable” (CNBC). Intermarché told AFP that affected stores would have to throw out all the products concerned. Managers who arrived on Monday morning met a wave of hot air from the freezers, according to the BBC, and BFMTV showed an ice cream freezer reading 26.5°C. Reuters reported that Paris resident Geert de Clercq watched staff at an Intermarché in the center of the city throw cartfuls of defrosted food into a giant trash container while melted ice cream spilled from broken tubs onto the street.

The Paris prosecutor’s office opened an inquiry that its cybercrime section is running with the national anti-cybercrime office, OFAC, to identify victims and tie them to the reported losses (Euronews, Sud Radio, both relaying AFP). It described a “large-scale” attack on connected refrigerators at several major retailers. Prime Minister Sébastien Lecornu has told ANSSI, France’s national cybersecurity agency, to investigate, a person in his office told Reuters.

Store counts differ by outlet and by day, and nearly all of them come from the companies. Reuters counted at least 350 supermarkets by Tuesday and called that a small fraction of France’s total (Reuters, as carried by Interaksyon). Adding up the chain figures, the BBC arrived at about 400 stores out of roughly 15,000 in France. France’s finance minister, Roland Lescure, told parliament on Tuesday that about 700 of the country’s 10,000 supermarkets were affected, according to CBC. The two tallies use different denominators, and the coverage does not reconcile them. The table collects what each chain, or press coverage of it, has reported.

Losses are harder to pin down. Les Mousquetaires told CNBC it expects losses of “several million euros” and is waiting for a court-appointed bailiff and insurance experts to produce firmer figures. Thierry Cotillard, Intermarché’s chief executive, told BFMTV the attack would cost tens of millions of euros nationwide in a single day, according to Reuters. At store level, BFMTV reported that two supermarkets in Alsace, an Intermarché and a Super U in Haguenau, lost about €270,000 of stock between them (BBC, Infosecurity Magazine). Intermarché also warned that some insurance contracts may not cover this type of incident, so part of the bill could stay with the stores.

The fires are the most disputed part of the story. The Paris prosecutor’s office said, as relayed by AFP, that there had also been départs de feu, or fire outbreaks, and Reuters reported that the attack led to food losses and fires. The BBC reports small fires in some places. Les Mousquetaires told CNBC that overheated equipment produced smoke in some of its stores but that no fires have been reported. No public report names a store, a number or a cause. ANSSI’s alert, posted by director general Vincent Strubel on 13 October, told operators to disconnect the affected controllers and flagged a fire-start risk alongside lost stock, according to Solutions Numériques.

The Target Is a Store-Wide Refrigeration Controller

Many headlines say smart fridges, which suggests a consumer appliance with a Wi-Fi chip. The affected equipment is closer to a thermostat with a web server attached: a store-wide controller that regulates cold cabinets and cold rooms, schedules defrost cycles and raises alarms. Reuters explains that supermarket cold aisles typically run on control panels that manage a whole store’s fridges and freezers and can be reached over the internet, which lets maintenance companies adjust them from anywhere. Danfoss, the Danish manufacturer, built the AK-SM and AK-SC series that French authorities have focused on. Its AK-SM 800A system manager, for instance, hosts a web interface where administrators monitor circuits, set temperature thresholds, review alarms and manage users, according to Industrial Cyber’s description of Claroty’s research.

The internet connection exists for a reason. Schelcher said Coopérative U links its fridges to the internet so that staff can monitor temperatures around the clock and catch an anomaly (CBC). That monitoring makes the exposure hard to remove, because disconnecting a controller also cuts the alert path that protects the stock. Les Mousquetaires disconnected about 100 further stores as a precaution, the BBC reports, which shows the trade-off in practice.

Danfoss told news agencies on Monday that the affected units were running outdated firmware or had reached end of life and no longer receive software or security updates (CBC). It has said the affected equipment included the AK-SM 800, declared end of life in 2021, and the AK-SC 355, phased out in 2016, along with current AK-SM 800A units running outdated firmware (Cooling Post). The history is awkward. Danfoss’s own phase-out notice for the AK-SC 355, which carries a 2016 end-of-life date, names the AK-SM 800 series as its replacement. The BBC adds that Danfoss described the hit models as older ones whose security software could not be updated against a risk the company had identified.

Danfoss also told Reuters it was aware of incidents affecting a limited number of its refrigeration management systems in France. Retailers count stores and Danfoss counts systems, and the coverage does not reconcile the two. The company may not be the only vendor involved: according to Reuters, ANSSI said that refrigeration systems from Copeland and from Dixell, a Copeland brand, were also hit (TechNadu). Reuters reports that Copeland acknowledged a vulnerability and said it is working with affected customers (Reuters, as carried by Interaksyon). A Leclerc store in the Tarn lost its cold chain on Tuesday and uses Dixell controllers, France 3 reported, as summarized by Solutions Numériques.

The mechanism matters as much as the brand. Joshua Steinman, chief executive of industrial security company Galvanick, told Industrial Cyber that “defrost isn’t an exploit; it’s functionality.” In his reading, the most damaging commands in industrial settings are often legitimate ones, so defenders need records of who reached a controller, what changed and whether anyone authorized it.

How the Attackers Got In: What Is Known and What Is Not

No one has publicly established the entry route. CERT-FR, the French government’s computer emergency response team, issued alert CERTFR-2026-ALE-012 on 12 October reporting numerous compromises of Danfoss AK-SM and AK-SC controllers and telling operators to disconnect them or filter access to them (The Hacker News). A person in Prime Minister Lecornu’s office told Reuters that ANSSI had identified a vulnerability in software supplied by Danfoss. That account names no CVE and rests on one anonymous source.

Public research names the leading suspects, though nobody has tied them to this incident. A CISA advisory published on 20 May 2025 covers CVE-2025-41450, an authentication bypass in the AK-SM 800A: instead of checking the configured admin password, the device accepts one generated from its own date and time. Danfoss fixed it in version R4.2. A Danfoss advisory dated 1 August 2025 covers two more flaws, a command injection that can lead to remote code execution (CVE-2025-41451) and an Nginx configuration injection (CVE-2025-41452), both fixed in version 4.3.1. Claroty’s Team82 published its technical write-up on 9 August 2026, according to Reuters, and, according to Bloomberg, found thousands of these controllers reachable from the internet (The Next Web). The fixes had been public for 14 to 17 months by the night of the attack.

Experts quoted by Reuters and CBC lean toward the simple explanation. Three cyber experts told Reuters the attack looked easy to carry out, and CBC reports that the experts it consulted believe unapplied Danfoss updates and interfaces left open to the internet made it possible, while noting that the exact mechanism is still unknown. Two cautions apply. Claroty’s work covered the current AK-SM 800A, while the AK-SM 800 and AK-SC 355 are the older units Danfoss says were among those hit. And exposure is not proof of compromise: SOCRadar, which counted Danfoss interfaces visible from France, said so directly (SOCRadar).

The exposure counts also vary with the scanner. Predicta Lab’s Baptiste Robert found 2,752 exposed Danfoss systems on Monday, 1,590 of them in France (BleepingComputer), and Reuters found at least 4,500 Danfoss interfaces on Shodan on Tuesday without being able to confirm how many ran vulnerable software. French scanning firm Onyphe told Reuters that only a small minority of exposed units appeared to run the latest software.

Other routes fit the same facts. Synacktiv, a French security firm, told Solutions Numériques that a recent audit found refrigeration systems reachable from the internet with no password at all, and that the geographic spread of the attacks resembles opportunistic scanning for exposed devices, though it said that does not confirm the method. Retailer statements also frame the incident as a supplier-side event: Les Mousquetaires told CNBC that the attack hit a service provider responsible for managing refrigeration systems in some of its stores. That would make a compromised maintenance channel as plausible as a scanned device.

Attribution is equally open. A source with knowledge of the probe told Reuters that investigators are puzzled by the absence of a hacktivist claim or a ransom demand. Ken Munro of Pen Test Partners called it a basic cyber hygiene issue, said amateurs could be responsible, and doubted that a state actor would hit only a subset of French fridges. He also read the missing ransom demand as a sign that the attackers did not grasp the commercial losses they could inflict (Reuters, as carried by Interaksyon).

Why the Patch Has No Clear Owner

Whatever the entry route turns out to be, the public statements already show a structural problem, because they point in different directions. Danfoss says the units hit were legacy products with no updates, or current products on outdated firmware, and that it has long advised customers to keep supported systems current and to migrate old ones. Its own cybersecurity guidance tells customers to keep its devices off the internet, put them behind a firewall or VPN, change default passwords and run the latest software.

Retailers place the incident elsewhere in the chain. Les Mousquetaires told Bloomberg that the systems were managed by Danfoss and that the supplier, not the retailer, should file a complaint with the authorities (The Next Web). E.Leclerc said a refrigeration equipment supplier was attacked (The Hacker News). France’s retail federation, the FCD, has publicly called on Danfoss for full transparency about the incident (Reuters).

The remote-access design divides the job three ways. The manufacturer publishes the fix, the retailer or its contractor has to install it, and the contractor often holds the remote access. When a controller retired in 2021 is still running in 2026, which of the three was supposed to notice?

Danfoss’s upgrade path is staged, according to The Hacker News: systems on 3.x move to 4.x first, and a controller on the 5.3.x package cannot be downgraded. Journal de l’économie cited update complexity and cost as obstacles to patching. Those frictions help explain how a fleet this size can run behind its vendor’s updates, and the delay carries a price. Ritesh Kotak, a Toronto-based cybersecurity analyst, told CBC that a published firmware update shows attackers exactly which vulnerability to target, which is why he stresses installing updates quickly.

What Nobody Has Confirmed

The coverage so far mixes confirmed facts with plausible inference. As of 14 October, five points remain open.

•        The entry route. Whether attackers used a known flaw, an exposed interface with weak or no authentication, or a compromised maintenance account, nobody has said. CERT-FR had published no indicators of compromise as of 13 October (The Hacker News).

•        The attacker. No group has claimed responsibility, and no ransom demand has been reported. Investigators are puzzled by both facts, according to a source who spoke to Reuters.

•        The fires. The prosecutor’s office mentions them, the BBC reports small fires in some places, and Les Mousquetaires reports smoke without fires. No store, number or cause has been published.

•        The spread beyond France. Félix Aimé, a threat analyst at Sekoia, told Reuters that Spain, Portugal, Poland, Germany and Belgium also appear affected. SOCRadar noted that retailers in those countries had not confirmed incidents (SOCRadar).

•        The Copeland and Dixell link. It rests on ANSSI’s statement as relayed by Reuters. Copeland acknowledged a vulnerability and said it is working with affected customers, but the coverage reviewed here gives no product-level detail.

Two other incidents surfaced in the coverage and should stay separate for now. Dutch retailer Ahold Delhaize took systems offline after a cybersecurity incident that began Sunday evening. The company said it saw no reason to link the incident to the French attacks, and the Dutch NCSC does not believe the two are connected (Reuters, CBC). Reuters also noted that the French attacks resemble disruptions at about a dozen US military commissaries in late August. Since then, Army CID’s cyber field office, the Air Force Office of Special Investigations and the Defense Criminal Investigative Agency have been investigating those outages, where some freezers went into defrost mode (Stars and Stripes). No cause or manufacturer has been named, no findings had been announced as of 30 September, and the Defense Commissary Agency says separate freezer problems at four European stores were unrelated to the US outages (Stars and Stripes).

The Regulatory Clock Is Already Running

Politicians moved quickly to frame the attack as a food-security issue. Lescure told parliament on Tuesday that there was no risk to the food supply, but he acknowledged that the attack goes to the heart of the country’s food security (CBC). European Commissioner for Agriculture and Food Christophe Hansen condemned it on X and said the “weaponization of food” has become a reality (CNBC). Food distribution has been a target before, CBC notes: JBS shut down operations in Canada, the US and Australia in 2021, and a 2025 attack on United Natural Foods led to shortages at Whole Foods.

The timing also makes this an early test for EU product-security law. Since 11 September 2026, manufacturers of products with digital elements have had to report actively exploited vulnerabilities and severe incidents under the Cyber Resilience Act, through the single reporting platform that ENISA operates. The rules call for an early warning within 24 hours of becoming aware, a fuller notification within 72 hours, and a final report afterward (European Commission). Law firm Jones Day reads the duty as covering products placed on the market before the Act’s main obligations start in December 2027, and notes fines of up to €15 million or 2.5% of worldwide turnover for non-compliance (Jones Day).

Whether any manufacturer has to file depends on facts that have not been published: whether a vulnerability in its product was actively exploited, and whether it has reasonable certainty of that. Danfoss declined further comment on Tuesday, and its security advisories page listed nothing about the attack as of 13 October, according to The Hacker News. How the duty applies to a model retired in 2021 is an open question that this case may force.

The operator side moves more slowly. The French bill transposing the NIS2 directive, whose sectors include food production, processing and wholesale distribution, was due for debate in the National Assembly on 7 October. The conference of presidents postponed it a day earlier, citing a busy agenda, and France missed the EU’s October 2024 transposition deadline (Next.ink). For now, the EU duty already in force reaches the manufacturers, while the operator-side rules in France are still waiting for a parliamentary slot.

Five Checks for Operators This Week

The public record already supports a short list of actions for anyone running connected refrigeration, whatever the final forensic report says.

1.     Inventory every refrigeration controller and remote-access gateway, including units a contractor installed. Record the model, firmware version and end-of-life status. Sophos’s Rafe Pilling advises identifying exposed and vulnerable systems, validating patch levels and assessing compensating controls (Industrial Cyber).

2.     Take management interfaces off the public internet, or limit them to trusted addresses through a VPN. That is the first instruction in CERT-FR’s alert, and it matches Danfoss’s own guidance. Keep the monitoring itself: Schelcher’s stores connect their fridges so that someone can watch temperatures around the clock, and a VPN preserves that without publishing the login page.

3.     Update supported units and plan the replacement of end-of-life ones. Ask your vendor or contractor to confirm in writing which firmware each store runs. If a unit may already be compromised, CERT-FR’s sequence is to restrict network access to service providers, factory reset, apply all fixes, change secrets, check accounts and reconfigure.

4.     Put the contractor’s access in the contract: who connects, from where, with what authentication and with what logs. Steinman’s point about evidence of who reached a controller and what changed only works if those logs exist.

5.     Add temperature monitoring that does not report through the controller. A separate sensor with its own alert path catches a controller that is compromised or misreporting. That is our recommendation rather than an official one, and it follows from Claroty’s demonstration, as Infosecurity Magazine described it, that a compromised AK-SM 800A could cut cooling while showing false temperatures.

The Question That Outlasts the Investigation

Investigators will probably name an entry route, and a CVE number may follow. The harder question is the one the retailers and the manufacturer are already answering differently: who was responsible for keeping that controller current and its login closed? Operators that can answer that for every cold room they run will be better placed when the next advisory lands, and those that cannot are leaving the answer to whoever scans the internet first.

The post Hackers Heated French Supermarket Fridges Up to 50°C. The Real Gap Is Who Maintains the Controllers appeared first on DataFLOQ.

Of course, it is convenient and sometimes remarkably useful. Still, speaking is a lot more than producing grammatically acceptable sentences.

Meanwhile, human conversation remains unpredictable, emotionally charged, and deeply contextual. For instance, a speaker may choose the correct words. Still, they might sound hesitant, dismissive, rehearsed, or strangely detached.

AI feedback can identify language patterns. However, human feedback often explains what those patterns actually mean inside a living conversation. That distinction matters more than it first appears.

1. Context Changes the Meaning of Good Speech

Experienced teachers understand that effective speaking depends on –

1. The situation
2. Relationship
3. Cultural setting
4. Purpose of the exchange.

For instance, iTalki tutors can notice when a learner sounds technically accurate but socially awkward. Then, they might suggest language that feels warmer, more natural, or more appropriate for a particular audience.

That kind of adjustment comes from human judgment. It is obviously not mere error detection.

Meanwhile, AI tools generally evaluate speech against recognisable linguistic patterns. They may flag –

• Filler words
• Repetition
• Pronunciation problems
• Unclear sentence structures.

Still, they can struggle when a speaker deliberately breaks a rule for humour, emphasis, diplomacy, or personal style. Usually, human listeners recognise that intention faster. More importantly, they will explain whether it worked.

2. Emotional Nuance Requires Human Interpretation

Speaking practice can expose insecurity. For instance, a learner may –

• Lower their voice after making a mistake
• Avoid difficult expressions
• Rush through an answer to prevent interruption.

Although AI can measure pace and volume, it may not understand why those changes occurred. Meanwhile, a human coach can connect the behaviour to –

1. Anxiety
2. Uncertainty
3. Fear of judgment.

As a result, the feedback becomes more precise. Instead of saying, “Speak more slowly,” a teacher might explain that the learner accelerates when discussing unfamiliar subjects. That observation identifies a trigger rather than treating the symptom. It also gives the learner something practical to watch during future conversations.

Human feedback also carries emotional weight. It might be a well-timed nod, a short pause, or a calm correction. These can reassure a learner without lowering expectations.

By contrast, constant automated scoring may turn speaking into a performance dashboard. Useful for tracking, perhaps. Not always useful for confidence.

3. Human and AI Feedback Solve Different Problems

The strongest distinction becomes clearer when both feedback types sit side by side. AI performs well when the task involves consistency, repetition, and measurable language features.

Humans become more valuable when meaning depends on interpretation, personality, or changing social conditions.

Therefore, treating human and AI feedback as direct substitutes misses the larger point. They operate at different levels. AI often describes visible features of speech, while a skilled person interprets the social effect those features produce.

4. Unscripted Interaction Exposes Hidden Weaknesses

Many learners speak comfortably when answering familiar prompts. However, real conversations rarely follow clean scripts.

People interrupt, change topics, misunderstand references, or respond with unexpected emotion. Human practice introduces this messiness naturally, and that is precisely why it remains valuable.

During live interaction, a teacher can test whether the learner can:

• Clarify a misunderstood point without becoming defensive.
• Adjust an explanation after noticing confusion.
• Re-enter a conversation after interruption or hesitation.

These abilities are difficult to reduce to a score. Moreover, they depend on awareness of another person’s reactions.

AI simulations can imitate disruption, but they usually operate within designed boundaries. Human responses remain less predictable, and therefore closer to the conditions learners eventually face.

5. Fluency Alone Cannot Judge Credibility

In most cases, professional speaking requires credibility rather than perfect fluency. For instance, the following issues must sound trustworthy:

• The manager presenting difficult news
• A candidate answering an interview question
• A specialist explaining a technical issue.

Of course, grammar matters. Nevertheless, audiences also judge –

1. Restraint
2. Confidence
3. Timing
4. Sincerity.

Human reviewers can detect when polished language begins to sound memorised. They may notice that an answer contains strong vocabulary but avoids the actual question.

Likewise, they can identify when excessive formality creates distance. AI may praise structural accuracy in the same response because the measurable components remain sound.

Furthermore, human feedback can challenge the speaker’s reasoning. A coach may ask whether an argument feels convincing, whether an example supports the claim, or whether the speaker has considered another interpretation.

At that point, speaking practice becomes thinking practice too. The value moves well beyond pronunciation.

6. Personalised Feedback Needs Memory and Judgment

AI systems can store progress data and adapt exercises to previous performance. Even so, meaningful personalisation requires more than remembering errors.

A human teacher can recognise that a learner communicates confidently about work but becomes guarded during informal conversation. That pattern may reflect –

• Personality
• Experience
• Cultural habit.

Accordingly, the teacher can decide when to correct or when to challenge. They can then judge when to let the conversation continue. Immediate correction sometimes helps. At other times, it destroys momentum.

Human instructors make that judgment in real time. They use subtle signals that automated systems may overlook or misread.

Better Speaking Emerges From a Hybrid Model

AI feedback works best as a practice multiplier. It offers repetition, availability, and fast detection of surface-level problems. Human feedback provides interpretation, adaptation, emotional awareness, and social realism. Neither contribution needs exaggeration.

For routine drills, automated tools can save time. They reveal patterns that learners might otherwise miss. However, human feedback still holds the stronger position when the goal involves –

1. Persuasion
2. Cultural sensitivity
3. Spontaneity
4. Trust
5. Genuine conversational confidence.

The practical model is hybrid. Use AI to practise frequently and track recurring issues. Then use human sessions to test whether improved speech actually lands well with another person.

Because in the end, successful speaking is not simply speech that a system can analyse. It is speech that another human being can understand, trust, and respond to.

The post The Human Advantage: Where AI Feedback Falls Short in Speaking Practice appeared first on DataFLOQ.

What Happened

Danish authorities disclosed on October 5 that hackers abused a private company’s legitimate access to the Central Person Register, or CPR, the database that assigns every resident, deceased person, and emigrant a unique identification number. The breach exposed names, addresses, and CPR numbers tied to roughly 8.75 million records, according to figures the attacker provided to Danish newspaper Politiken. That total includes about 5.5 million living people and 3.3 million deceased ones, out of the roughly 11 million records the CPR holds. Denmark’s actual population is about 6 million.

The compromised account belonged to Pays ApS, a small Odense-based data services firm that helps business customers verify addresses, deaths, and relocations through CPR lookups. Managing director Sophie Laursen confirmed on October 9 that an external party had misused the company’s authorized access. Between September 10 and September 20, the intruder ran roughly 14 million automated searches, 8.75 million of which returned usable records. Nobody caught it in real time. A ministry staffer discovered the breach on October 2 after noticing an unusually large invoice tied to Pays’s search volume, 12 days after the activity had already stopped.

Politiken later reported that at least three Pays accounts, including an administrator account, used the password “123456.” Separately, a person claiming responsibility told the paper, through an encrypted chat conducted in English, that they got in using a leaked password belonging to a former Pays employee. Danish authorities have confirmed neither account of how the attacker got in, nor the attacker’s identity. Digitalization Minister Christina Egelund called the incident “extremely serious” and ordered a review of the CPR system’s security controls. Datatilsynet, Denmark’s data protection authority, and the National Special Crime Unit are both investigating.

The attacker also told Politiken that artificial intelligence wrote the extraction program, a script more than 40,000 characters long, saying “any AI model launched within the past year or so would be able to do it.” Two experts who reviewed the code for Politiken agreed it looked AI-generated. The clearest evidence: a bug sent the program searching for people born in 1877 instead of 1977, so the leaked dataset is missing an entire birth year by accident.

Why It Matters

None of the attacker’s claims, about the password, the AI tooling, or the stated intent not to sell the data, have been independently verified by Danish authorities. But the pattern underneath them is familiar. CPR numbers function in Denmark the way Social Security numbers function in the United States: as both a universal identifier and, in practice, a stand-in for proof of identity. Banks, employers, hospitals, and landlords all request it. Once a number like that leaks at population scale, treating it as a secret stops making sense, because tens of millions of other parties already hold it too. Danish cybersecurity specialist Jan Kaastrup told broadcaster TV 2 that the CPR number’s role as an implicit secret is now simply “broken.”

That is also why the discovery method matters as much as the breach itself. Fourteen million automated queries ran over ten days through one vendor account, and nothing in the CPR system’s monitoring flagged it as abnormal. An invoice surfaced the intrusion, not a security alert. The pattern echoes other 2026 incidents, including the Thomson Reuters court-records exposure and a string of vendor-credential breaches across healthcare and law firms, where the attacker broke nothing technical. They simply used access that was already there.

The AI detail is the part generating headlines, and it deserves scrutiny rather than alarm. A 40,000-character script that enumerates sequential ID numbers and pulls matching records is not a sophisticated exploit. What AI assistance plausibly did was compress the time and skill needed to write that script from days to minutes, not invent a new attack method. The century bug, mistaking 1977 for 1877, is a two-digit-year error that has dogged software since long before large language models existed, apparently inherited now by the tools that write software today.

Our Take: Our take is that the password and the AI detail are the least important facts in this story, even though they are the most quotable. The real failure is structural: a government handed broad, standing search access to a small vendor with no apparent limit on query volume, no real-time anomaly detection, and no requirement for multi-factor protection on an account that could enumerate an entire nation’s population. Fixing the password would not have fixed that. Denmark’s advice to lean on MitID and one-time codes instead of the CPR number is the right move, but it is also an admission that the number itself is now permanently compromised for roughly 8.8 million people.

Expect more governments to face this same reckoning as AI tooling keeps lowering the cost of mass-scale enumeration attacks against databases built around a single, static identifier. The institutions that get ahead of it will be the ones that stop treating identity numbers as secrets and start treating vendor access itself as the perimeter worth watching.

The post Denmark’s Hackers Didn’t Need Skill, They Needed an AI and a Bad Password appeared first on DataFLOQ.

What Ahold Delhaize Has Confirmed

The disruption began on Sunday afternoon, October 11. In a statement published Monday morning, Albert Heijn said it, Etos and Gall & Gall were dealing with a cybersecurity incident and had taken immediate measures to protect their IT environment, including the temporary, precautionary shutdown of several systems. Websites, apps and customer service went offline. Logistics systems were disabled, so online grocery orders cannot be delivered until further notice. As of Monday evening, the company had not said when services would return.

Inside the stores, the picture is patchy. Checkouts and this week’s national Bonus offers work. The Mijn AH app, the personal Bonus card, Koopzegels and Air Miles do not, and electronic shelf labels may still show last week’s promotional prices, which can sit below the price at the till. Albert Heijn has asked customers who miss personal discounts to keep their receipts so it can compensate them later.

The problem reaches beyond the Netherlands. Parent company Ahold Delhaize said it is “experiencing a security issue within its network” and is taking some systems offline “to help protect them,” according to BleepingComputer. Grocery Dive reported that the websites of all five U.S. banners, including Food Lion, Hannaford and Stop & Shop, were down Monday morning. In Belgium, Delhaize customers could not place new orders for the week, VRT reported, although stores in Belgium and Luxembourg stayed open.

The Damage Sits in the Warehouse

The more consequential detail came Monday afternoon. Versteden-van Duijn told NOS that work had stopped completely in some distribution centers because “we had to switch the systems off preventively.” Albert Heijn runs six distribution centers in the Netherlands, and she would not say which ones are still working. Trade outlet Warehouse Totaal, citing RTL Z and an email sent to suppliers, reported that only the automated center in Zaandam was still running and that suppliers had been asked to deliver there within agreed time slots. Albert Heijn has not confirmed that account.

Fresh food is the most exposed. Albert Heijn expects frozen goods, bread and long-life products to stay better stocked than fruit and vegetables, and says availability will vary by store. “Of course shelves will run empty if this lasts longer,” the CEO said. One delivery driver told NOS that all drivers were thrown out of the delivery app at the same time and could not log back in.

The stoppage likely reflects a precaution rather than a confirmed hit on the warehouses. By the CEO’s account, the centers stopped because Albert Heijn turned its systems off, and the company has not said attackers reached them. It has also not explained why checkouts survived while logistics did not. One reading is that a store can keep selling what is already on its shelves, while a distribution center with its systems switched off cannot ship anything. That remains an inference, because the company has not described how its systems are connected.

No Confirmed Data Breach, and No Quick Answer

No one has said data was stolen. Albert Heijn says protecting the data of customers, employees and partners is its priority, and NOS reported that the company cannot yet say whether outsiders entered its IT systems; the CEO said the investigation is still running. Ahold Delhaize calls the event a cybersecurity incident without naming an attacker, a method or a cause, and its spokesperson did not answer whether any devices were encrypted, BleepingComputer reported. NL Times relayed RTL’s reading that the signs point to a cybercriminal attack. Neither company has confirmed that.

The group’s own history argues against expecting a fast answer on data. Ahold Delhaize disclosed a cyberattack on its U.S. banners in November 2024, which halted Hannaford’s online orders for several days. It confirmed in April 2025, about five months later, that files had been stolen, as the INC Ransom group claimed up to 6 terabytes of data from the U.S. division. Grocery Dive reports that the company later said the incident potentially exposed sensitive information of more than 2 million people.

Why the Shutdown Costs So Much

Shutting systems down early carries a clear defense. “The short-term pain of a planned shutdown is almost always preferable to an uncontrolled one,” said Muhammad Yahya Patel, a cybersecurity advisor at Huntress, in comments published by Cybernews. Marks & Spencer shows how large the bill can grow when an outage drags on. After an April ransomware attack in 2025, M&S stopped accepting online orders. In May it told investors to expect online disruption through June and into July, and it put the hit at around £300 million for 2025/26 before mitigation, BleepingComputer reported, adding that food sales also suffered from reduced availability. Albert Heijn has not said it faces ransomware, and the M&S figures are a reference point only.

Our Take: a grocer’s first-week exposure is availability. A stolen database takes months to size, as 2024 showed. A stalled distribution center shows up on shelves within days, and the products Albert Heijn flagged first, fruit and vegetables, are perishable and turn over quickly. Retail incident plans should put the restart of logistics at the center, alongside the recovery of data.

Two answers will decide how this ends: whether anyone outside the company reached its systems, and how long some of its distribution centers stay shut. If 2024 is any guide, the first will take months to settle. The second will show up in the produce aisle much sooner.

The post Albert Heijn’s Cybersecurity Incident: Checkouts Work, but Some Distribution Centers Don’t appeared first on DataFLOQ.

Venture Money Is Moving Behind the Liquor Store Counter

On October 6, 2026, New York start-up Santé announced a $15 million Series A led by FINTOP. CEO Darren Fike took a part-time job at a TriBeCa wine store to learn the trade, and he told AlleyWatch he noticed that his favorite New York wine shops ran on legacy systems that held them back. The company sells wine and liquor retailers an operating system that covers receiving and ordering, e-commerce, delivery app management and marketing, with AI agents that read distributor invoices and price slow-moving inventory. Santé says nearly 1,000 independent retail locations use the platform, together processing more than $2 billion in annual gross merchandise volume, and that its business grew 500% in the past year. Those figures come from the company.

Denver-based Scotch attacks the same market from the hardware side. Crunchbase News reported in June 2026 that Scotch raised a $20 million Series A led by VMG Partners. Its product combines point-of-sale hardware, payment processing and a back-office suite that handles state-by-state regulatory compliance. Scotch reports more than $1 billion in processed payment volume and growth above 500% year over year. The company was incorporated in January 2024 and employs about 45 people. Its CTO, Dan Chen, previously served as chief architect at Drizly.

Retailers buy from distributors, and that step has its own software. In 2022, Provi and SevenFifty began combining their data into a single B2B marketplace for beverage alcohol, listing more than 1,200 distributor portfolios and over 700,000 products with pricing, Bar & Restaurant reported.

Drizly supplies the backstory for Scotch’s CTO. Uber paid $1.1 billion for the alcohol delivery app in 2021, then announced in January 2024 that it would close the business in March, saying it wanted to focus on Uber Eats. Two funding rounds do not make a market, but together they suggest investors now favor the software that stores run on over the apps that sit between a buyer and a bottle.

Discovery Apps Turned Drinkers Into Data

Vivino and Untappd took the consumer side first. Vivino, which dates to 2010, reports more than 70 million users and a database of more than 19 million wines. Untappd, owned by software developer Next Glass, widened its scope in November 2025 from beer to spirits, wine, canned cocktails, THC-infused drinks, coffee and soft drinks, according to The Drinks Business. The app has logged more than 1.5 billion beer check-ins since 2010, at more than a million bars, restaurants, breweries and retail outlets, and its drink categories have grown from 450 to more than 900.

Both apps center on the product in the glass. A rating tells a brand or a retailer what people liked. It says much less about whether the person who poured the drink knew what they were doing, and venue-level scoring fills that space.

A Vodka Brand Builds a Scorecard for Bars

Vodka Polar Bear comes at the data question from the venue side. The brand, based in Hilversum in the Netherlands, sells a grain vodka that its production notes describe as distilled five times, filtered slowly near freezing and bottled at 37.5% ABV in Schiedam, the traditional Dutch distilling city. It ships in 70 cl and 1 liter bottles, in cases of six, to liquor stores and the hospitality trade, and it dates its history to 1985. On VodkaPolarBear.com, the company also runs Vodka Lab, a searchable knowledge hub, next to a library of 22 cocktail recipes.

The Cocktail Index page states its purpose in plain terms. Translated from Dutch, it says you do not drink a cocktail every day, so when you treat yourself, you should do it properly. In late September 2026, the brand published the method behind the index and its companion Mocktail Index. Each venue receives a score from 0 to 100, built from four parts: offering (30 points), craftsmanship (25), guest experiences (35) and consistency (10). A score requires a confirmed menu or other proof of the offering, plus independent, drink-focused guest experiences. The index also labels how strong the available evidence is: strong, sufficient or limited. When information runs thin, the score stays blank. The index treats a blank as unknown rather than zero.

The first published scores show the method at work. Super Lyan, the bar at the Kimpton De Witt hotel in Amsterdam, earned 92 out of 100: 29 of 30 for offering, 24 of 25 for craftsmanship, 31 of 35 for experiences and 8 of 10 for consistency. Rosalia’s Menagerie Cocktailbar, also in Amsterdam, scored 82. Super Lyan’s profile also shows the blank-score rule in practice. A visitor praised its mocktails, but the index says it does not yet have enough separate experiences to score them.

The index is built for venues in the Netherlands and other European countries. Volume one lists 21 Dutch addresses, most of them inside hotels, and two carry scores so far. The other 19 show the label “score volgt,” or score to follow. The venue profiles and the full method are public at VodkaPolarBear.com.

Why a Published Method Matters

The weights carry a view of what makes a good night out. Guest experiences hold the largest share at 35 points, ahead of the offering at 30, so a strong menu cannot carry a score alone. Evidence labels add a second signal: the index rates the evidence behind Super Lyan’s cocktail score as strong and the evidence behind Rosalia’s as sufficient.

The implication is that the label may matter as much as the number. A score with a stated confidence level tells readers how far to trust it, which a bare figure cannot do. Operators can read the four components as a checklist of their own: offering, craftsmanship, guest experiences and consistency.

The Mocktail Index also sits in a category with real money behind it. Crunchbase News reported in June 2025 that Athletic Brewing had raised close to $250 million in equity, and that global sales of non-alcoholic beer rose 9% by volume in 2024, per IWSR.

Credibility remains the open question for any new index. Volume one holds two scores, and the methodology article does not yet say how often scores refresh or how venues are chosen. Those details will matter as the list grows, because readers trust a published method most when they can watch it applied consistently across many venues.

The Next Test for Drinks Tech

Wine and beer already have apps that tell drinkers what to expect from a bottle or a pint. The Cocktail Index applies the same logic to the bar and shows its working while it does. If Vodka Polar Bear keeps adding scored venues, the question for the rest of drinks tech becomes who else will publish a method this openly.

The post Drinks Start-Ups Digitized the Bottle Shop. Vodka Polar Bear’s Cocktail Index Takes On the Bar appeared first on DataFLOQ.