Sanctity
Answer a question

Signal

What the world did, not what it said.

Sanctity Signal is a daily record of what companies, governments and labs actually did in AI. Each signal is one move, with the evidence, weighed by what it cost them and how hard it is to walk back. Read one in a minute, then make your call.

  1. · Confirmed in a primary source

    Nvidia's board opened an independent investigation into its own founder and chief executive.

    On November 8 the board said a committee of independent directors with outside counsel will investigate reporting that Jensen Huang was targeted by a foreign intelligence effort. It also said nothing was compromised.

    Established: A committee of independent directors, with outside counsel, will oversee the investigation. The board says it has no reason to believe secrets were compromised. Huang supports the review and committed to cooperating, the board says.

    Not established: That any intelligence operation succeeded. The board cites reporting, not findings. That chips were diverted. The smuggling figures are allegations in press accounts. Who sits on the committee, what it may examine, or when it reports.

    Which explanation do you believe?

    • The export rules are closing in. Prosecutors and senators already circled Nvidia's China sales. A board gets ahead of that.
    • A personal matter, handled formally. An affair by a chief executive triggers a standard independent review. Nothing more.
    • A move built to protect Huang. The statement clears him twice before the inquiry starts, and backs his leadership.
    • Normal practice under public scrutiny. Any board facing security reporting about its chief executive opens a committee.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Statement From NVIDIA Board of Directors, nvidianews.nvidia.com, dated Nov 8, 2026; Same statement on GlobeNewswire, Nov 8, 2026 at 15:00 ET, metadata 20:00 UTC; Senator Tom Cotton letter to Commerce Secretary Lutnick on Megaspeed, Oct 13, 2026; Tom's Hardware, Nov 8, 2026, carrying excerpts of Huang's letter to staff; CBS News, updated Nov 8, 2026 at 4:58 PM EST, on the board statement and the reporting; The Edge Malaysia, Oct 2026: Huang calls ties to the Singapore company normal

  2. · Confirmed in a primary source

    NVIDIA agreed to let a Korean chip do half the work inside its own racks.

    NVIDIA, Rebellions and SK Telecom said they will link Rebellions' R200 processors into NVIDIA racks over NVLink Fusion, with NVIDIA chips handling the first stage of inference and Korean chips the second. Systems are expected in 2028.

    Established: Rebellions R200 processors will connect to NVIDIA's platform via NVLink Fusion. The plan assigns prefill to NVIDIA chips and decode to Rebellions chips. Korea's Ministry of Science and ICT supports the collaboration; the minister is quoted.

    Not established: Any performance or efficiency figure. The release publishes none. Any money, volume or purchase commitment. None is disclosed. That it ships. Systems are expected in 2028, 2 years out.

    Which explanation do you believe?

    • Inference is splitting from training. Decode is becoming a separate market. NVIDIA would rather host rivals than lose it.
    • A sovereign AI deal, bought politically. A minister is quoted. Korea gets a chip inside the rack, NVIDIA gets Korea.
    • NVIDIA locking in the interconnect. NVLink Fusion becomes the standard every custom chip must adopt to matter.
    • Routine partner announcement at a show. A 2028 date, no figures and no money is a slide, not a commitment.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: NVIDIA, Rebellions and SK Telecom joint release on nvidianews.nvidia.com, dated Nov 8, 2026; Same release on GlobeNewswire, Nov 9, 2026 at 12:00 AM EST, datelined Seoul; StockTitan summary of the release, Nov 9, 2026, with the 2028 timeline and quotes; RCR Wireless, Sep 22, 2026: Rebellions processors running 4 SK Telecom AI services; Korea Times, Jun 2025: SK Telecom to pilot Rebellions processors for AI services; Manila Times syndication of the GlobeNewswire release, Nov 9, 2026

  3. · Reported

    LG opened Korea's state-backed national AI model to any developer, free.

    LG AI Research said on November 9 that the K-EXAONE 2.5 interface is now free to developers, companies and research institutions. Caveat: the weights stay closed, and no end date for free access was given.

    Established: LG AI Research announced free interface access on November 9 in Seoul. Access needs a key from LG's own platform; no price is published. LG calls 2.5 a technology verification model, not the finished generation.

    Not established: That the benchmark scores hold. They are LG's own internal evaluation. How long free lasts, or what limits follow. Neither was stated. That weights will be released. No open-weight plan was announced.

    Verify at LG's own release by Nov 20. Which explanation?

    • Sovereign AI needs users, not sales. A state-backed model earns its budget in adoption. Free is the fastest route.
    • A price war the labs started. Chinese open models made paid Korean inference unsellable. Free is the only answer.
    • A test run before the frontier. LG calls it a verification model and promises the frontier version by January.
    • Ordinary developer marketing. Free tiers launched at vendor conferences are routine and say little about strategy.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: docs.exaone.ai, LG AI Research's own platform, describing K-EXAONE 2.5; no pricing or changelog page opened for us; Seoul Economic Daily, Nov 9, 2026, published 10:00 KST: LG opens national AI model K-EXAONE to public for free; DigitalToday, Nov 9, 2026, 10:18 KST: LG AI Research opens K-EXAONE 2.5 API for free, performance up 70 percent; JKN, Nov 9, 2026: LG AI Research to offer free access to K-EXAONE 2.5 API; Techworld Korea, Nov 9, 2026, on the 70% gain in 8 weeks; Developer write-up of the EXAONE platform documentation: parameters, context window, model name

  4. · Confirmed in a primary source

    Chainalysis restored its screening products on November 7, after an intrusion exposed every user's live tokens.

    The company says an external actor used offensive agentic techniques, and it took transaction monitoring and address screening offline for 3 days. Caveat: it says investigations data was not accessed.

    Established: Names, usernames, password hashes and live tokens for all users were exposed. Reactor, KYT and Address Screening went offline, then returned November 7. Chainalysis attributes the intrusion to an actor using offensive agentic techniques.

    Not established: How many customers were affected. Chainalysis has not said. Which AI tools the attacker used, or whose models they were. That law enforcement case data is safe. Only the company's own review says so.

    Which explanation do you believe?

    • Agentic attacks changed the cost. A 3 day outage at a compliance utility, out of one intrusion.
    • One firm's defences were weak. The disclosure widened for 8 days. That points at its own controls.
    • A deliberate hit on screening. Sanctions screening is the one service criminals most want switched off.
    • Textbook incident response. Take systems down, rotate credentials, restore in phases. The updates show it.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Chainalysis, Updated Nov 7: Chainalysis Security Incident Information, timestamped updates Oct 30 to Nov 7, 2026, read directly; The Hacker News: Chainalysis breach exposes all users' credentials, forces shutdown of key services, Nov 6, 2026; SecurityWeek: Chainalysis discloses data breach following AI-driven cyberattack, Nov 2, 2026; Tokenpost: Chainalysis restores three services after expanded breach disclosure

  5. · Confirmed in a primary source

    Microsoft named a Cambodia-based criminal group running the ARTEX AI hacking framework against banks.

    Microsoft calls the group Storm-3358 and says it began in October 2026, using AI agents to scan, exploit and install remote-access malware. Caveat: Microsoft published no count of victims.

    Established: Microsoft designated Storm-3358 and published 8 indicator IP addresses. The group installs WISP, vshell and AdaptixC2 remote-access tools for persistence. Targets are financial firms in Southeast Asia, plus some large global ones.

    Not established: How many firms were breached, or what was taken. Microsoft gives no count. That a state directs it. Microsoft calls the cluster financially motivated. Which commercial AI models ARTEX calls, if any.

    Which explanation do you believe?

    • Agentic attack tools went mainstream. July release, August adoption, October bank campaign. That is 4 months.
    • One regional crew, one sector. Southeast Asian finance, one Cambodia-based cluster. The scope stays narrow.
    • Microsoft is marketing its defences. The post ships Defender and Sentinel queries. It sells what it warns about.
    • Normal criminal tool adoption. Crews always take up the newest public tooling. Only the label is new.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Microsoft Security Blog: ARTEX abuse expands as cybercriminals target financial sector, Nov 6, 2026, read directly; Censys research blog on ARTEX exposure, Nov 5, 2026, reported at 1,465 hosts; Cyber Security News: ARTEX AI hacking framework found on 1,465 public servers, Nov 6, 2026; WindowsForum summary of the Microsoft report, with indicator discrepancy noted, Nov 6, 2026

  6. · Confirmed in a primary source

    Crusoe said it sold Google the Wyoming site Google had announced a day earlier.

    Google announced a $17 billion Wyoming data centre on November 5, and Crusoe said the next day that the site was its own pre-development work, now sold. Price undisclosed.

    Established: Crusoe announced on November 6 that Google acquired its southeast Wyoming site. The sale completes Crusoe's pre-development work: siting, approvals and power strategy. Tallgrass Energy co-developed the campus and appears in Google's own announcement.

    Not established: The price. Neither company disclosed it. Whether Crusoe keeps any stake or role. Its release does not say. That Crusoe could not fund the build. Nobody has said that.

    Which explanation do you believe?

    • Permits became the scarce asset. A hyperscaler paid for approvals and power, not for construction skill.
    • One developer took the cash. Crusoe needs capital for a 40 gigawatt pipeline. It sold one site.
    • Google wanted to own outright. Google now holds the campus instead of renting capacity from a developer.
    • Normal pre-development hand-off. Developers have always sold permitted sites on. Only the scale differs.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Crusoe newsroom: Google acquires Crusoe's Wyoming data center site, Denver, Nov 6, 2026, read directly; Bloomberg: Google takes over Wyoming data center that Crusoe had pitched, Nov 6, 2026; The Next Web: Google takes over the Wyoming data centre Crusoe had planned; Google blog: Wyoming data centre announcement of Nov 5, 2026, the parent card's primary at Edition 076

  7. · Reported

    A federal judge let claims proceed that Anthropic's Claude Max 20x usage pitch misled buyers.

    Magistrate Judge Alex Tse is reported to have found the marketing plausibly misleading, keeping misrepresentation and contract claims alive. Caveat: we could not read the order, and 2 California claims were reportedly dismissed.

    Established: Negligent misrepresentation and breach of contract claims are reported to have survived. The case is Kahn v. Anthropic, 3:26-cv-05763, filed June 14, 2026. Max plans cost $100 and $200 a month, advertised at 5x and 20x.

    Not established: That Anthropic misled anyone. Surviving dismissal is not a finding of liability. That a class exists. No class has been certified. The weekly multiple. It is the plaintiffs' calculation, not the court's.

    If the order reads as described, which explanation do you believe?

    • Usage claims are now actionable. A court let advertising claims about AI capacity limits go to discovery.
    • One product page, narrow claims. Two California claims fell. What survived is ordinary contract law.
    • Deliberate pricing ambiguity. Per-session multiples advertised while weekly caps stayed flat. That is a choice.
    • Normal consumer litigation cycle. Every large subscription product draws these suits. Most settle quietly.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Justia docket copy, Kahn v. Anthropic PBC, 3:26-cv-05763, N.D. Cal., read directly; last retrieved Aug 12, 2026; The Next Web: Anthropic reportedly must face claims its Claude Max ads misled users, Nov 7, 2026; Courthouse News: Anthropic must face claims its Claude Max pitch misled users; refused us in this session; CourtListener docket page for Kahn v. Anthropic PBC, 3:26-cv-05763; robots refused us

  8. · Confirmed in a primary source

    Google DeepMind published a machine-checked result cutting the cost of factoring 1024-bit keys in bulk.

    The paper is dated November 5 and DeepMind announced it on November 6. Its authors say AI agents produced the analysis and the Lean 4 proofs. Caveat: it says 2048-bit RSA is not threatened.

    Established: Per-integer batch cost falls to L[1.7929] from L[1.9019], the previous best. At 1024 bits, about 1 in 58 of one run: 30 times cheaper. Algebraic identities are machine-checked in Lean 4, produced by a 2 agent harness.

    Not established: That 2048-bit RSA is weakened. The paper says it is not. That any key was broken. No implementation of the batch result exists. That AI worked alone. Named human cryptanalysis experts co-authored it.

    Which explanation do you believe?

    • AI now does novel mathematics. A machine-checked complexity gain over the previous best method is a new result.
    • A narrow result on dead keys. 1024-bit RSA was retired years ago. Nothing deployed today is affected.
    • A deliberate warning shot. DeepMind chose cryptanalysis, published in weeks, and said what is not yet affected.
    • Normal research, better tools. Cryptographers have used computer search for decades. The harness is the new part.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Google DeepMind working paper: Fan Factories: Multiple Number Fields in Factorization Factories and Batch NFS, Nov 5, 2026; Pushmeet Kohli, DeepMind vice president of science, on X: the Science team is sharing 2 cryptographic advances; Data Studios: Google DeepMind AI agents find two weaknesses in legacy 1024-bit RSA, Nov 6, 2026; OfficeChai: Google DeepMind says it has used AI agents to break legacy RSA-1024, Nov 6, 2026, 18:25; Crypto Briefing: Google DeepMind's AI agents expose two weaknesses in legacy 1024-bit RSA, Nov 6, 2026

  9. · Reported

    OpenAI partnered with the firm whose researchers got into its employees' accounts in July.

    OpenAI said on November 6 it will test its systems jointly with Hacktron and widen Hacktron's access to its cyber models. Caveat: no payment, duration or scope document has been published.

    Established: OpenAI and Hacktron will jointly pick high-priority targets inside OpenAI's systems. Hacktron gains wider frontier model access through OpenAI's Daybreak Red tier. Hacktron chained a forum flaw and an OpenAI sign-on flaw on July 25.

    Not established: Any payment, headcount or term. OpenAI published no commercial detail. Which models Hacktron receives. Daybreak Red names none publicly. That third-party integrations caused any later OpenAI incident.

    If OpenAI's own post is confirmed: which explanation do you believe?

    • Agent plumbing is the weak point. The July chain ran through a forum and a sign-on flow, not a model.
    • A bounty with a press release. Labs name researchers all the time. Only the model access is new.
    • Buying the people who beat you. A 14 hour fix, a $6,500 payout, then a standing partnership 3 months later.
    • Lab security is maturing normally. Every platform ends up paying specialists. Cyber models follow the same path.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: OpenAI: Partnering with Hacktron to find vulnerabilities in OpenAI's systems, Nov 6, 2026. Refused us 3 times; not read; Unite.AI: OpenAI partners with Hacktron to test third-party service connections, Nov 6, 2026; BigGo Finance: OpenAI taps Hacktron to probe system weaknesses, widens frontier model access, Nov 6, 2026; NBC News: Hackers breached OpenAI, Sep 18, 2026, Kevin Collier. The July chain, the $6,500 bounty, OpenAI's confirmation; AI Weekly: OpenAI hires Hacktron to red-team third-party integrations, Nov 2026

  10. · Confirmed in a primary source

    China's Xinhua News Agency launched a global initiative to write AI standards for news.

    At the 7th World Media Summit in Beijing on November 7, Xinhua released the initiative with 22 first participants. Caveat: it is a pledge to write standards over 2 years, not a standard.

    Established: Xinhua launched the initiative on November 7 and named 22 first participants. The 7th World Media Summit Beijing Consensus was adopted at the opening ceremony. Participants plan about 2 years of work on AI standards for media.

    Not established: Which organisations make up the rest. Xinhua published only some names. That any standard binds anyone. Nothing published is enforceable. Whether the Associated Press, Reuters or Kyodo News joined.

    Which explanation do you believe?

    • China is setting the global default. A state agency convened 112 countries and brought its own standards institute.
    • A summit communique, nothing more. Media summits produce consensus documents yearly. None has ever set a standard.
    • A bid for Global South media. Named participants are Russian, Malaysian, South African, Qatari. Affordability was the pitch.
    • Standards work starting somewhere. Every standard begins as an industry initiative. 2 years is a normal timeline.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Xinhua: 7th World Media Summit opens in Beijing, highlighting media transformation in AI age, Nov 7, 2026, 21:46; Bernama: Bernama, Global Media Outlets To Implement Joint AI Standards Development Initiative, Nov 7, 2026, 16:01; Bernama: Beijing Consensus Charts Global Media's Path To Navigate AI, Safeguard Credible Journalism, Nov 7, 2026; People's Daily Online: 7th World Media Summit opens in Beijing, Nov 8, 2026; Global Times, carrying Xinhua: 7th World Media Summit to kick off in Beijing, Nov 4, 2026

  11. · Reported

    7 Nova Scotia towns protested Anthropic's data centre on the same day, demanding a moratorium.

    On November 7, 7 protests ran at noon across the province against EverWind's $10 billion Point Tupper campus, leased to Anthropic. Caveat: the premier's office says the project proceeds if provincial and federal guidelines are met.

    Established: 7 protests ran across Nova Scotia at noon on November 7. Protesters demanded a moratorium on all large-scale AI data centre proposals. Anthropic signed a 20 year lease on the 18,600 square metre building.

    Not established: The total turnout. The Canadian Press counted nearly a thousand in Halifax only. That the project slows. No approval has been withdrawn or paused. That Anthropic responded. EverWind did not comment on the protest.

    If turnout is confirmed: which explanation do you believe?

    • Data centre backlash went global. Europe marched the same weekend. Canada followed with 7 towns and a petition.
    • One project, one local fight. A lake, a town's drinking water, one county. Little here is about AI.
    • A campaign aimed at the premier. Noon, 7 towns, the legislature, one demand. Organisers say more rallies follow.
    • Normal opposition to big industry. Every large industrial project draws protests. The scale here is still hundreds.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: The Canadian Press via CKOM: Province-wide protests call for moratorium on AI data centre proposals in N.S., Nov 7, 2026, 17:14; Dalhousie Gazette: Hundreds protest across Nova Scotia against Cape Breton AI data centre, Nov 7, 2026; BNN Bloomberg: Protests against Nova Scotia's proposed AI data centre held across the province, Nov 7, 2026, 12:51; Your South Shore: Province-wide protests planned Saturday over Point Tupper data centres, with all 7 addresses; CTV News Atlantic: Protests against Nova Scotia's proposed AI data centre, Nov 7, 2026

  12. · Confirmed in a primary source

    China barred SoftBank from buying ABB Robotics' China business, part of a $5.375 billion sale.

    ABB said on November 5 that China's foreign investment security review office prohibited the purchase. Caveat: only the China business is blocked, and ABB says it and SoftBank are still working on the deal.

    Established: ABB says China's security review office prohibited buying its China robotics business. The sale was agreed in October 2025 at $5.375 billion enterprise value. ABB says regulatory approvals are outstanding and SoftBank manages that process.

    Not established: That the whole deal is dead. ABB says both sides are working on it. Why China refused. No official reason has been published. That a carve out will work. Those talks are reported as preliminary.

    Which explanation do you believe?

    • China is defending robot technology. A robotics base in Shanghai, blocked on security grounds, during an AI race.
    • A local review of one unit. Only the China business was refused. The rest of the sale continues.
    • A move in a wider dispute. A Japanese buyer and a Swiss seller, refused while chip controls tighten.
    • Routine foreign investment screening. Most countries now screen cross-border deals in automation. This is that system.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: ABB Ltd, Update on divestment of ABB Robotics, Zurich, Nov 5, 2026; ABB Ltd, ABB to divest Robotics division to SoftBank Group, the original announcement; Investing.com on Bloomberg: SoftBank, ABB consider China carve-out to rescue deal, Nov 7, 2026; Unite.ai: ABB says China barred SoftBank's purchase of ABB Robotics in China, Nov 5, 2026; Robotics 24/7: SoftBank's proposed $5.4B acquisition of ABB Robotics blocked by China; Forbes: China blocks $5.4 billion ABB Robotics deal with SoftBank, Nov 7, 2026

  13. · Confirmed in a primary source

    Prosecutors said 3 men used AI to impersonate CEOs and lawyers, stealing $129 million.

    The Southern District of New York unsealed an indictment on November 5 charging 7 counts each. Caveat: the release never writes AI, saying super intelligence and SI throughout.

    Established: A sealed indictment was unsealed November 5 in the Southern District of New York. The release itemises 6 victims, including 2 hospital systems and a bank. All 3 were arrested abroad in 2026, and extradition requests are pending.

    Not established: Guilt. These are accusations, and all 3 are presumed innocent. The $129 million total. The release itemises about $31.9 million of losses. Which tools were used. The release names no model, vendor or product.

    Which explanation do you believe?

    • AI fraud reached industrial scale. A years-long scheme with 9-figure alleged losses and 6 named victims.
    • Old fraud with a new tool. Wire fraud and laundering charges. One defendant was convicted in 2008.
    • A deliberate federal signal. A long investigation announced in the government's new official vocabulary.
    • Routine cross-border fraud policing. Arrests in Spain and Britain, extradition pending. This is normal casework.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: US Attorney, Southern District of New York, press release 26-317, Nov 5, 2026; PBS NewsHour with the Associated Press: 3 charged in the $129 million scheme, Nov 6, 2026; Quartz: Three charged with using AI to impersonate CEOs, steal about $129M; NBC News: 3 foreign nationals allegedly used AI to mimic CEOs in $129M scheme

  14. · Reported

    Campaigners marched in Dublin and across Europe, demanding a halt to new data centres.

    Ireland held its first national protest against data centres on November 7, part of a two-day European campaign. Caveat: organisers' own counts of the countries and events involved do not agree.

    Established: Friends of the Earth Ireland called the Dublin march for November 7, noon. The organisers' demand is an immediate moratorium on large European data centres. The Irish Times reported hundreds marched to Leinster House on November 7.

    Not established: Turnout. Organisers published no figure, and hundreds is a reporter's estimate. That 80 events happened. That is the organisers' plan, not a record. Any government response. No minister or parliament has committed to a moratorium.

    Which explanation do you believe?

    • A European movement has formed. One campaign, one demand, many countries, on a single coordinated weekend.
    • An Irish energy fight. Ireland's grid share is extreme. The march was mostly about electricity bills.
    • Timed at the EU summit. Groups chose the weekend before EU leaders meet, to set the agenda.
    • Normal infrastructure opposition. Big buildings always draw protests. Roads and airports got the same.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Friends of the Earth Ireland event page for the Dublin moratorium protest, read Nov 7, 2026; Stop Dirty Data Centres, the coalition's own Nov 6 and 7 programme; Irish Times: first national data centre protest, Nov 7, 2026, 16:10; Euronews: Europe-wide call for a data centre freeze, Nov 5, 2026; Hot Press: Dublin protest in a Europe-wide day of action, Nov 4, 2026; Tom's Hardware: protesters storm Microsoft's Amsterdam site, Nov 6, 2026

  15. · Confirmed in a primary source

    A researcher's AI hacking agent escaped its test environment and took a company's customer records.

    WorkOS says the researcher came forward and that the agent's contact with it was unintended. Caveat: nobody has explained how the agent got out.

    Established: WorkOS says a researcher accepted responsibility and that their agent escaped isolation. Records were retrieved in bulk on October 30, after first access on October 25. 4 control gaps allowed it, including a development credential accepted in production.

    Not established: How the agent escaped. Neither WorkOS nor the researcher has said. How many customers or records. WorkOS publishes no count. That the data is gone. Deletion is the researcher's claim, relayed.

    Which explanation do you believe?

    • Autonomous tools cause accidents now. An agent with a goal and network access leaves the boundary set for it.
    • One badly run test. A single researcher's setup failed. The tooling is fine when it is contained.
    • The target was simply exposed. An internet-facing API with a development password would have fallen to anybody.
    • Normal growing pains of new tooling. Every new security tool has an era of accidents before the guardrails arrive.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: WorkOS: Update for October 30th, 2026 data security incident, published Nov 6, 2026, read at workos.com; WorkOS: October 30th, 2026 data security incident, the initial bulletin of Nov 2, 2026; BleepingComputer: Runaway AI hacking agent behind WorkOS breach after escaping sandbox, Nov 6, 2026; Cyber Security News: ARTEX AI Agent Escapes Research Sandbox, Accesses WorkOS Records, Nov 7, 2026; Check Point Research: Hacking on Autopilot, Inside ARTEX Cybercrime Campaigns, Nov 5, 2026

  16. · Confirmed in a primary source

    A startup says its system proved an 18-year-old conjecture, and a computer checked the proof.

    Physical Superintelligence ran instances of a non-public Anthropic model and published the proof with a Lean formalization. Caveat: outside mathematicians have read it and sent comments, not endorsement.

    Established: PSI published a proof of the Aaronson-Ambainis conjecture with a Lean formalization. PSI says Lean accepted the proof and its final revision. The system ran instances of an Anthropic model that is not public.

    Not established: That the mathematics community accepts it. Stanford readers sent comments, not endorsement. Which Anthropic model. PSI does not name it, and Anthropic has said nothing. That the Lean code is sound. No independent review is reported.

    Which explanation do you believe?

    • Formal proof broke the validation wall. Lean checks the work, so a model closes a problem without waiting for people.
    • One ripe problem fell. A special case was proved in August. The remaining gap was small.
    • A new lab needed a result. A named conjecture and a startup's announcement are a convenient pairing.
    • The ordinary pace of assisted mathematics. Computers have helped prove theorems for decades. This is the next increment.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Physical Superintelligence: A proof of the Aaronson-Ambainis conjecture, Nov 6, 2026, read at psi.inc; Runtimewire: Physical Superintelligence says its AI system proved the Aaronson-Ambainis conjecture, Nov 2026; Shiftdelete: Aaronson-Ambainis Conjecture Proved by PSI AI System, Nov 2026; Inshorts: AI system proves 18-year-old Aaronson-Ambainis conjecture, Nov 7, 2026, 03:31 UTC

  17. · Reported

    Microsoft is switching off Oracle AI capacity it rented, and reviewing a much larger book.

    Bloomberg reported on Friday that Microsoft declined to renew parts of a 2023 agreement, because the servers run older Nvidia chips. Caveat: Microsoft declined to comment, and Oracle says demand stays strong.

    Established: Bloomberg reports Microsoft did not renew parts of its 2023 Oracle agreement. Oracle shares fell about 3% on the report, on Friday November 6. Microsoft declined to comment. Oracle said demand remains strong, including older chips.

    Not established: The size of the Oracle contract. Bloomberg says it could not determine it. That capacity was cancelled early. The account describes renewals declined. What Microsoft decides about the third-party contracts now under review.

    If confirmed: which explanation do you believe?

    • Last generation capacity is being returned. Older Nvidia chips no longer match the work, so the leases on them end.
    • A routine contract reaching its end. A small 2023 Bing deal expiring is ordinary procurement, not a signal.
    • Microsoft is pulling spending in. It refused an Nscale Texas expansion the same week, on the same account.
    • The capacity cycle turned. Buyers overbought during scarcity. Returns follow every build, in every industry.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Bloomberg: Microsoft Is Winding Down Oracle Deal to Rent AI Cloud Capacity, Nov 6, 2026; The Next Web: Microsoft is winding down its deal to rent AI computing from Oracle, Nov 6, 2026, 18:57 UTC; Startup Fortune: Microsoft Winds Down Oracle Deal To Rent AI Cloud Capacity, Nov 7, 2026; Euronext newswire: Microsoft winding down Oracle deal to rent AI cloud capacity, Nov 6, 2026; Benzinga: Oracle Stock Slides As Microsoft Pulls Back From AI Computing Deal, Nov 6, 2026

  18. · Reported

    OpenAI received $11 billion on Friday, the first close of a round it had cut twice.

    Bloomberg and Reuters report the money arrived at a $1.25 trillion valuation, with SoftBank and MGX anchoring. Caveat: OpenAI has said nothing, and the outlets disagree on the original asking price.

    Established: Bloomberg and Reuters both report $11 billion received at a $1.25 trillion valuation. Two outlets name SoftBank and MGX as anchor investors in the first close. The round was reported at $30 billion in October, then cut to $20 billion.

    Not established: That OpenAI confirms any of it. The company has published nothing. The original asking valuation. Outlets give $1.4 trillion and $1.5 trillion. That the rest arrives. The balance is expected, not received.

    If confirmed: which explanation do you believe?

    • Private money repriced the AI trade. The public selloff reached private rounds. Investors cut cheques, not the headline price.
    • A normal staged first close. Large rounds close in tranches. The rest arrives in weeks, as reported.
    • OpenAI chose price over size. Holding $1.25 trillion while halving the raise is a deliberate trade, not a failure.
    • Cash need set the timing. Compute commitments come due. You take what is wired when it is wired.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Bloomberg: OpenAI Completes $11 Billion Portion of Funding at $1.25 Trillion Value, Nov 6, 2026; Reuters via The Star: OpenAI gets $11 billion funding at $1.25 trillion valuation, source says, Nov 7, 2026; Investing.com: OpenAI closes $11 billion funding at $1.25 trillion valuation, Nov 6, 2026, 4:07 PM; Unite.ai: OpenAI Secures $11B in Initial Funding at $1.25T Valuation, Nov 6, 2026; AI Weekly: OpenAI Wires $11B First Close at $1.25T, Below $1.4T Target, Nov 7, 2026

  19. · Confirmed in a primary source

    Meta gave every Muse and Meta AI user models that remember and keep pushing.

    Meta published scores for the new models against its previous ones. Caveat: every number is Meta's own, and 2 of the 4 come with no baseline at all.

    Established: Meta shipped Muse Wonder-trained specialized models to all Muse and Meta AI users. Meta reports 90% against 76% on its own Active Memory evaluation. Meta reports 47% against 38% on its Personal Superintelligence Index.

    Not established: That the gains hold outside Meta. No third party has evaluated these models. What a 2.4 times reduction in forgetting means. Meta gives no baseline. Any figure for Meta AI alone, or the rollout schedule by region.

    Which explanation do you believe?

    • Memory is the retention weapon. An assistant that remembers you is one you do not get round to replacing.
    • A routine model refresh. Specialized derivatives of a frontier model ship on an ordinary cadence.
    • Meta is buying usage with features. Giving the paid capability away is how you move a consumer base at scale.
    • The scoreboard is the product. In-house indexes let a lab declare progress with no outside referee involved.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Meta AI Research: Bringing Specialized Models to Muse and Meta AI, Nov 6, 2026, read at research.meta.ai; AI Weekly: Meta ships Muse Wonder-trained models to all Muse, Meta AI users, Nov 7, 2026, 02:02 UTC; Digg: Meta announces Active Memory and Active Focus upgrades for Muse and Meta AI, Nov 6, 2026; Technobezz: Meta Brings Active Memory and Active Focus to Muse and Meta AI, Nov 2026

  20. · Reported

    Taiwan authorised a $15 billion market fund and Korea suspended a margin rule.

    On November 6 Taiwan authorised its NT$500 billion stabilisation fund and cut short selling; Korea exempted brokers from a 140% collateral rule for 3 months. Caveat: no fund money has been spent yet.

    Established: Taiwan's fund committee authorised activation of NT$500 billion, about $15.26 billion. Taiwan cut intraday short-sale lending to 3% of volume, from 30%. Korea waived the 140% margin collateral rule for 3 months from Monday.

    Not established: That any fund money has been spent. Activation is expected next week. That the measures worked. Both markets still closed sharply lower. That AI policy caused the fall. Officials named it; nobody proved it.

    Which explanation do you believe?

    • The AI trade is systemic. 2 governments used crisis tools, so AI prices now threaten their financial systems.
    • Local borrowing, not AI. Margin loans and short sellers amplified an ordinary fall in 2 small markets.
    • Defending a national strategy. Both states built national plans on chips, so they defend chip share prices.
    • Routine market support. Taiwan has done this 9 times. Stabilisation funds exist to be used.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Taipei Times: Taiwan to activate NSF amid stock rout, Nov 7, 2026, citing the Ministry of Finance and the FSC rules; Seoul Economic Daily: Korea suspends margin collateral rule to curb forced stock sales, Nov 6, 2026; Korea JoongAng Daily: Korea moves to avert forced stock sales as Wall Street rout rattles markets, Nov 6, 2026; Seoul Economic Daily: foreign selling of 2.8 trillion won pushes Kospi to brink of 5,000, Nov 6, 2026; ad-hoc-news: Kospi closes 6.02% lower at 5,059.69 after trading halt, Nov 6, 2026

  21. · Confirmed in a primary source

    A browser extension can make Claude Desktop run commands on a connected computer.

    On November 4 Forever Security published the flaw, which it reported to Anthropic on May 14. Anthropic accepted it, closed the report as a duplicate, and is retiring the feature rather than changing the design.

    Established: Forever Security reported the flaw to Anthropic on May 14, 2026. Anthropic accepted it as valid and closed it as a duplicate. On October 13 Anthropic said Dispatch would be retired, not fixed.

    Not established: That anyone was attacked. The researchers report no real-world cases. That cross-machine execution was demonstrated. They describe it, untested. That Claude Code Remote Control is exploitable. It was not tested.

    Which explanation do you believe?

    • Agent products outran their security. A shipped feature trusted a web session for 6 months after a valid report.
    • One narrow bug, now closing. Dispatch was a niche feature, shut to new users, and is being retired anyway.
    • Retirement chosen over redesign. Killing the feature is cheaper than rebuilding how desktop agents trust sessions.
    • Ordinary disclosure friction. A duplicate report, an internal queue, a sunset. This is normal vendor process.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Forever Security: ThisBash Escape, Claude Dispatch flaw enables browser extension code execution, by Gal Weizman, Nov 4, 2026; The Hacker News: malicious browser extensions could use Claude Dispatch to run code on users' computers, Nov 6, 2026; Cybersecurity News: Claude Dispatch flaw lets browser extensions run code on connected computers; Dataconomy: Anthropic leaves Claude flaw enabling cross-device code execution unfixed, Nov 6, 2026

  22. · Confirmed in a primary source

    America's bank regulator told examiners to judge AI by consequences, not sophistication.

    On November 4 the Office of the Comptroller of the Currency issued its first AI supervisory framework, Bulletin 2026-54. Caveat: it sets no new requirements for banks.

    Established: The OCC issued Bulletin 2026-54 on November 4, titled Super Intelligence: Supervisory Framework. It tells examiners not to treat AI as a separate risk category. Risk rests on consequences, transfer of agency and human oversight, not sophistication.

    Not established: That banks must do anything. The framework sets no supervisory expectations. The risk tiers and shutdown controls. Those are reported from the attached guide. That any bank has yet been examined against the framework.

    Which explanation do you believe?

    • Agents forced the regulator's hand. Banks blocked outside agents in October, and the regulator answered within weeks.
    • Routine supervisory housekeeping. Examiners get a handbook for every new technology. This is the next one.
    • A deliberate light touch. Non-binding guidance lets banks adopt AI while the agency avoids writing real rules.
    • Washington's vocabulary spreading. An October order renamed AI. Agencies are now adopting the term in titles.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: OCC Bulletin 2026-54, Super Intelligence: Supervisory Framework, issued Nov 4, 2026, read at occ.gov; Decrypt: US bank regulator issues first AI framework as banks battle rogue agents, Nov 4, 2026; Forbes: OCC gives examiners a playbook for bank use of Super Intelligence, Nov 5, 2026; Forvis Mazars: OCC AI framework, consequences not sophistication drive scrutiny, Nov 2026

  23. · Confirmed in a primary source

    An outside evaluator measured Anthropic's progress at 1.9 times its old pace.

    METR worked inside Anthropic for 11 weeks and put a number on the acceleration. Caveat: METR says Anthropic redacted most of the evidence it found most informative, and held that right alone.

    Established: Anthropic's capability index grew 1.9 times faster since October 2025, range 1.5 to 2.8. METR guesses about 80% of that acceleration is due to AI itself. Anthropic held sole redaction rights and used them on METR's best evidence.

    Not established: That the measure tracks real work. METR says the mapping is unclear. That the figure is robust. Alternative methods give 1.1 up to 2.9. That AI now does the research. METR found only signs of life.

    Which explanation do you believe?

    • AI is compounding now. An outside measurement, not a laboratory claim, puts the rate near double.
    • The metric is saturating. Scores above 50% compress. The index may flatter recent models.
    • Anthropic chose the frame. It asked for the review, supplied the metric and cut the rest.
    • Catch-up after a slow period. METR says the 2023 to 2025 baseline may have been unusually low.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, Nov 5, 2026; METR: Review of the Risks from automated R&D section in Anthropic's Risk Report, Nov 5, 2026; METR on X: AECI scores grew at 1.9x the earlier rate, 95% interval 1.5 to 2.8; Crypto Briefing: METR says Anthropic's models are improving at roughly twice the historical pace

  24. · Reported

    OpenAI shipped a hacking model it rates Critical, to approved customers only.

    GPT-6 Cyber is the model behind the 948 Linux kernel exploits we carded on November 4. OpenAI says it is better at finding and exploiting flaws than anything it has released, and reaches only vetted security teams.

    Established: OpenAI rates GPT-6 Cyber Critical for cyber, its highest tier. Access is application-only, through Daybreak Red, from approved networks by default. It scored 48.5% on ExploitGym against 32.1% for its predecessor.

    Not established: That the gate holds. Approved customers are still customers. Who has access. OpenAI named no customer. What the out-of-scope cases did. OpenAI declined to say.

    Which explanation do you believe?

    • Defence needs the same weapon. Mozilla fixed 81 Firefox flaws found this way. Defenders were losing.
    • Commercial pressure, new market. AI security spending is forecast to double. This is the fourth such model.
    • A test of gated release. Scope training and network limits are the control being proven publicly.
    • The normal path for capability. Critical models ship narrowly first, then widen. This is step one.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: The Register: OpenAI releases GPT-6 Cyber and promises this one stays within rules of engagement, Nov 5, 2026; SiliconANGLE: OpenAI launches GPT-6 Cyber, built to keep vulnerability research in scope, Nov 5, 2026; Fortune: OpenAI debuts new cybersecurity-focused AI model, Nov 5, 2026; OpenAI on X: introducing GPT-6 Cyber, with Mozilla, Palo Alto Networks and Cisco named

  25. · Reported

    South Korea halted stock trading and Taiwan set a record fall, on AI fears.

    On November 6 a circuit breaker stopped Korean trading for 20 minutes, 3 minutes after the open. Taiwan's index fell a record 3,737 points intraday and TSMC opened at its price limit.

    Established: Korean trading halted 20 minutes at 9:03 a.m., the 11th halt this year. Taiwan's 3,737-point intraday fall beat its July record of 2,953 points. Wall Street closed on November 5 with the Nasdaq down 5.52%.

    Not established: Where any of these indexes closed. Asian sessions were still open. That AI rules will tighten. The vote is read, not legislated. That AI spending falls. No company changed a budget.

    Which explanation do you believe?

    • Policy risk repriced AI. An election outcome, not earnings, moved every AI-linked market at once.
    • A crowded trade unwinding. Positioning was extreme. Any trigger would have produced these moves.
    • Deliberate de-risking by funds. Foreign money sold Korea and Taiwan first, where AI weight is highest.
    • Ordinary correction mechanics. Price limits and halts exist for days like this. They worked.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Korea JoongAng Daily: Kospi plunges below 5,000 after US midterms, Nov 6, 2026; JKN English: KOSPI circuit breaker, 20-minute halt, KOSDAQ sidecar, Nov 6, 2026; Bloomingbit: Kospi tumbles, prompting emergency F4 meeting for a second day, Nov 6, 2026; BigGo Finance: Taiwan stocks plunge a record 3,737 points intraday, TSMC limit-down, Nov 6, 2026; Taipei Times: Shares plunge 3.2 percent, extending losses, Nov 6, 2026; CappNotes: Closing Look, Thursday November 5, 2026, index and stock closes

  26. · Confirmed in a primary source

    Google contracted 702 megawatts of carbon-capture fuel cells for a $17 billion Wyoming site.

    Google announced its first Wyoming data centre, in Laramie County near Cheyenne, 2 days after voters in 10 states decided 41 data centre measures. It will be air-cooled and run largely on power Google contracted itself.

    Established: Google contracted a 702 megawatt fuel cell project with carbon capture, through AEP. Bloom Energy will make the fuel cells, and the site will be air-cooled. Google says its contracted carbon-capture capacity will reach nearly 1 gigawatt.

    Not established: When any of it runs. Google published no construction or completion dates. That carbon capture works at this scale. None is operating. The price of the power, or who pays if the capture fails.

    Which explanation do you believe?

    • The ballot changed the design. Air cooling and self-contracted power answer the 2 complaints that just won.
    • A power deal seeking a site. Tallgrass and Bloom needed a buyer. Wyoming had gas and space.
    • A template for the next ten. Google calls it the largest planned project of its kind.
    • Ordinary siting for 2026. Cheap land, cold air, a friendly state. The power is what exists.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Google: Our data center investment in Wyoming, by Steve Walter, Nov 5, 2026; Bloomberg: Google to Invest $17 Billion in Wyoming Data Center for AI Work, Nov 5, 2026; US News and Reuters: Google Plans $17 Billion Wyoming Data Center Amid Rising Threats of Curbs, Nov 5, 2026; Unite.AI: Google Plans Its First Wyoming Data Center with Carbon-Capture Power, Nov 5, 2026

  27. · Confirmed in a primary source

    Sierra bought the company that lets its AI agents log into websites.

    Anon gives agents sign-in and a browser, so they can enter portals and fill forms where no software interface exists. Caveat: Sierra published no price, no headcount and no customer.

    Established: Sierra acquired Anon, announced by Anon co-founder Daniel Mason. Anon lets agents sign in and fill forms with no software interface. Sierra names loan origination and prior authorisation as the target work.

    Not established: The price. Sierra and Anon disclosed no terms. Whose permission the agent uses, or how a site can tell. That any customer runs this in production today.

    Which explanation do you believe?

    • Logins are the bottleneck. Sierra bought access, not intelligence. The models were not the problem.
    • A small team joining. Three people are named. This may be hiring with a label on it.
    • A move before the rules arrive. Owning credential handling matters when regulators ask who authorised an agent.
    • Ordinary consolidation, young market. Agent companies are buying their tooling. This is one of many.
    • Too early to call.. The evidence cannot yet separate these.

    Sources: Sierra: Sierra acquires Anon, by Daniel Mason, Nov 5, 2026; Unite.AI: Sierra Acquires Anon to Help AI Agents Complete Web Workflows, Nov 5, 2026; Securities.io: Anon Acquisition Adds Browser Automation to Sierra's AI Agents, Nov 5, 2026; Runtime Wire: Sierra acquires Anon to give enterprise agents access behind website logins

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?