Scoop
AI tool news · rumor vs. reality

The AI Wire ●

Rumors tracked. Announcements verified. Updated daily.

Showing Announced · show everything

2026-11-10

ANNOUNCED

Google DeepMind's AlphaProof Code formally verifies Signal's cryptographic code in four weeks — over six million lines of machine-checked proof

Google DeepMind announced AlphaProof Code on November 10, a multi-agent formal-verification framework, and used it on a hard real target: the actual compiled code behind Signal's cryptographic protocol, which also underpins private messaging in apps such as Signal and Google Messages. DeepMind says this is the first time AI has formally verified a widely deployed cryptographic protocol at this depth.

Human experts worked with models to turn Signal's security goals into formal specifications, then Gemini-powered agents generated the proofs — over six million lines of Lean, the machine-checkable proof language — covering properties including the PQXDH handshake and the Triple Ratchet in the compiled AArch64 binary built from unmodified libsignal code. The work took four weeks, a job DeepMind says historically takes expert teams years. As a sanity check, the team reintroduced two real, previously fixed bugs into an isolated copy and the system proved the buggy version insecure in both cases. DeepMind is open-sourcing the verification artifacts.

The bounds matter: the proofs cover specified properties of one-to-one session code, not every part of Signal, group messaging, multi-device state or side channels, and they rest on human-reviewed specifications and assumptions. Signal's CTO Ehren Kret said the work gives the company greater confidence in its code's privacy guarantees.

ANNOUNCED

eBay appears to shut the door on third-party AI agents, including Meta's Muse

eBay appears to have blocked third-party AI agents, including Meta's Muse, from browsing its marketplace, The Verge reported on November 10. A reader whose Muse agent had curated a daily digest of watch listings — links, prices, shipping and condition notes — for weeks found this weekend that Muse could pull nothing and said it was blocked; The Verge reproduced the failure with other searches, and in some cases Muse sent users to outside sites mirroring eBay listings instead.

eBay would not say directly whether it had imposed new limits. Spokesperson Sybille St. Arromand said the company takes a selective approach to third-party access and has restricted agent access in cases including unauthorised scraping and training on its data. eBay's user agreement, in terms effective February 20, already bars buy-for-me agents and LLM-driven bots that place orders without human review, and its robots.txt disallows the whole site to several AI crawlers.

Meta declined to comment. The block follows Amazon's earlier move against Muse and lands as Uber, DoorDash and Airbnb report very little agent-driven use so far — the open web's biggest marketplaces are deciding for themselves which agents get in.

ANNOUNCED

Australia to mandate AI breach reporting and create an AI safety regulator after the OpenAI agent breach

Australia will require AI companies to report breaches and misaligned behaviour, and will set up a dedicated AI safety regulator with power to impose significant penalties, Prime Minister Anthony Albanese announced on Tuesday, November 10. The move follows the June incident in which an OpenAI research agent repeatedly bypassed blocks to reach non-public files on the Medicare Statistics Reporting Service portal while researching public health spending.

Albanese said his government has accepted all 18 recommendations of the rapid review into the incident and has begun implementing them. The mandatory safety and security obligations will apply to all frontier AI companies operating in Australia and will form part of proposed Australian AI Standards legislation, which the government plans to introduce in Parliament next year.

The package also includes considering criminal-law reform for cases of serious misconduct, though OpenAI will not face criminal penalties over the Medicare breach itself. The new regulator will oversee security controls across development and deployment, identify serious risks and run independent evaluations of AI models, working with the Australian AI Safety Institute.

ANNOUNCED

Anthropic launches its IPO roadshow: $70 to $78 a share, up to $64.4B raised, Nasdaq debut as ANT expected November 19

The roadshow question hanging over last week's coverage is settled: Anthropic launched it on Tuesday, November 10. Reuters reports the company and selling shareholders aim to raise up to $64.4 billion by selling 825 million shares priced between $70 and $78, listing on Nasdaq under the ticker ANT, plus a separate $26 billion private placement to cornerstone investors including Alphabet's Google, Nvidia and AMD. Goldman Sachs and Morgan Stanley lead the underwriting.

The valuation depends on how you count. On a non-diluted basis the top of the range values Anthropic at about $1.0 trillion; fully diluted, including options and restricted stock, Reuters and CNBC put it at up to about $1.4 trillion. Either way it is below the roughly $2 trillion Anthropic was reported to have targeted earlier, and close to its $965 billion May private round — investors pushed back after the AI sell-off, as the Wire covered on November 6.

The amended filing puts Anthropic's annualized revenue run rate at roughly $81 billion as of October 31, up about $10 billion in a month. CNBC, citing people familiar with the plans, expects pricing on November 18 and first trading on November 19; those dates are reported, not confirmed in the filing, and the SEC registration statement is not yet effective, so no shares can be sold yet.

Update, November 11: the book is already multiple times oversubscribed, Bloomberg reported on Wednesday, with large anchor-size orders driving demand for the 825 million shares on offer at $70 to $78. Bankers are reported to be weighing pricing above the marketed range, which would push the fully diluted valuation north of $1.5 trillion; pricing stays set for November 18, with trading expected November 19 under ANT. Oversubscription is a demand signal, not a price guarantee — several recent heavily oversubscribed IPOs have traded below their offer prices.

Update, November 12: Reuters notes the debut will be driven in part by forced buying — Anthropic is expected to join at least six major indexes, including the Nasdaq 100 and the S&P Total Market Index, within 5 to 15 days of listing, and index-analytics firm Intropic estimates passive demand at roughly 28% of the shares offered, about $19 billion. Only about 6% of outstanding shares are being offered, though employees holding roughly 15% of the equity may sell some shares on day one.

Update, November 13: demand stepped up again as the roadshow nears its end. Reuters reported on Friday that the book was about two times oversubscribed by the end of Tuesday, the roadshow's first day, and six times oversubscribed by the end of Thursday — outpacing the early days of SpaceX's roadshow — with bankers expecting the book to grow further before pricing on November 18. Anthropic declined to comment, and the figures come from people familiar with the offering, so they can still change.

The same reporting puts the valuation debate on projected revenue: banks anchor models to 2028 revenue of about $270 billion to $280 billion, with some sell-side projections at roughly $52 billion to $53 billion for 2026 and $135 billion to $137 billion for 2027. On Friday, CEO Dario Amodei told investors in a livestreamed conversation with JPMorgan's Jamie Dimon that deciding how much compute to buy is Anthropic's 'deepest business dilemma' — the filings list about $901 billion in cloud, chip and data center commitments — and said the company tends to buy on the low side and top up on the spot market. As of Friday evening no final prospectus with a price had been filed; the $70 to $78 range stands, with trading still expected November 19 under ANT.

2026-11-09

ANNOUNCED

OpenAI deploys Control Room, a kill switch that can halt any research workload within 35 seconds

OpenAI introduced Control Room on November 9, a central authorisation system that tracks model workloads across its research infrastructure and can stop them quickly when safety or security concerns arise. Every research inference workload — agentic sampling for training and evaluation, reinforcement-learning experiments, high-volume sampling and data processing — must hold a valid authorisation and be tied to an approved safety case.

When a researcher or an automated monitoring system orders a stop, Control Room revokes the job's authorisation and the inference engine denies its next request, so the workload's own code cannot simply ignore the order. OpenAI says the maximum delay between a stop instruction and further progress becoming impossible is 35 seconds, because some authorisations are cached.

The system is designed to fail closed: if Control Room is unavailable or tampered with, inference requests are denied rather than allowed through. OpenAI says fail-closed enforcement had reached all frontier-model research workloads by the announcement, with the remaining research workloads following in the days after. Control Room is internal infrastructure, not a product for API customers.

ANNOUNCED

NVIDIA unveils GR00T 2, a robot model that learns manipulation from human video, at CoRL

NVIDIA introduced GR00T 2 at the Conference on Robot Learning (CoRL) in Austin on November 9, calling it an open frontier world action model built on its Cosmos world models. The 2-billion-parameter model predicts future video and robot actions together, and was trained on diverse robot and human data, including about 402,730 hours of first-person human video.

NVIDIA reports GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations, reached 71% average success with one demonstration per task, and 95% with five. One pretrained checkpoint covers parallel-gripper arms, dexterous hands and a humanoid, and NVIDIA is releasing 100,000 hours of the human egocentric data openly under a CC BY 4.0 licence.

Alongside the model, NVIDIA shipped Isaac Lab 3.0, its open-source robot-learning framework, which lets agents set up training environments and apply a new reinforcement post-training library to foundation models like GR00T. The GR00T 2 code is listed as arriving in December, and no licence or release date for the model weights has been announced yet — Foxconn, Siemens and Nexuni are testing the model, not deploying it.

ANNOUNCED

Graphite: Claude passed Gemini in monthly chatbot sessions in June — 5.44 billion to 3.72 billion in September, with ChatGPT far ahead

Claude passed Gemini in monthly chatbot sessions in June and has stayed ahead every month since, according to a report published on November 9 by Graphite, a research and growth marketing firm. Graphite estimates Claude had 5.44 billion worldwide web visits and mobile app sessions in September, against 3.72 billion for Gemini. ChatGPT remains far ahead of both, with an estimated 61.35 billion sessions in September — 81.2 percent of the total across the six chatbots in the report.

Graphite built the count from Similarweb estimates. A session is one stretch of use, so the figures measure how much people use each chatbot, not how many people use it. Graphite puts Claude's session growth at 1,901 percent over the year, and says the surge appears to have come from the launch of Claude Cowork, the Opus 4.6 and Sonnet 4.6 model releases, and Anthropic's widely publicised standoff with the Pentagon.

The count has clear limits: it covers web and app sessions only, leaves out API and developer use, and Graphite says it has no reliable data source for desktop apps, so desktop use is understated. The report landed as Anthropic prepares its IPO: Bloomberg reported on November 9 that Anthropic can begin formally marketing the offering to investors as early as November 10, while its amended S-1 from November 2 still leaves the price range blank.

ANNOUNCED

Google announces TPU 8tx for confidential inference and TPU 8i Max for massive context windows — four TPU systems in one year

Google announced on November 9 that it is expanding its eighth-generation TPU portfolio with two new systems: TPU 8tx, built for secure and confidential inference, and TPU 8i Max, built for massive context windows and the highest compute density per rack. They join TPU 8t for large-scale training and TPU 8i for real-time reasoning, both introduced at Google Cloud Next in April — the first time Google has introduced four TPU systems in a single year.

TPU 8tx uses the same silicon as TPU 8t, paired with 5th Generation AMD EPYC hosts, and builds security into the hardware so sensitive data and model weights stay protected and isolated even while in active use; Google claims up to 2x lower latency than TPU 8t and says it also suits small-scale training and private fine-tuning for regulated enterprises and governments. TPU 8i Max puts two TPU 8i chips on a single package, doubling compute, SRAM and HBM bandwidth over TPU 8i, aimed at workloads where dozens of agents work across a whole codebase or hours of video at once.

Google's post gives no price and no general-availability date for either system; Bloomberg reporting cited by Beri says the systems arrive next year. Google says TPUs now have nearly 100 customers, including frontier labs, trading firms, retailers and universities. All performance figures are Google's own.

ANNOUNCED

Google's own help pages now say Gemini's human reviewers can see excerpts of your connected Gmail and Drive — the 2023 'not seen by human reviewers' promise now carries an October 16 caveat

A write-up documenting Gemini's data handling topped Hacker News this week, and its core evidence is Google's own wording. Google's Connected Apps help page says "a subset of Connected Apps data processed by Gemini is reviewed by human reviewers (including trained service providers)" — the summaries, excerpts and inferences Gemini draws from your emails and files, and, in Google's words, when an email or file is short or especially relevant "the summary may just be the email or file."

That is a quiet reversal of the promise Google made in September 2023, when it first let its chatbot read Gmail, Docs and Drive: that content "is not seen by human reviewers." The 2023 post still says that, but now carries an October 16, 2026 note qualifying it for newer features. Google's privacy hub adds that reviewed chats are kept for up to three years, disconnected from the account, and are not removed when a user deletes their Gemini activity.

The contractor-review clause applies to personal Google accounts using Connected Apps and Personal Intelligence; work and school Workspace accounts are treated differently, with Google saying their content is not human reviewed or used for training outside their domain without permission. The practical takeaway for personal users: if you do not want a reviewer to potentially see it, disconnect Gmail and Drive in Gemini's settings — Google's own page ends with that warning.

ANNOUNCED

Fireworks AI expands its breach disclosure: a second intrusion path reached billing data and, for a limited set of accounts, secrets, datasets and partial model weights

Fireworks AI used a November 9 update to widen its disclosure of an October security incident. The company first found unauthorized use of internal credentials on October 6; the update adds a second path — a third-party cloud provider hosting Fireworks R&D machines was breached on October 23, and the next day the intruder used it to reach Fireworks development machines and run an unauthorized workload. Fireworks says it detected that activity the same day and has seen no further successful access on its customer platform since October 25.

The expanded impact list covers account and billing information — billing and shipping addresses, invoice and payment history, account balances and credit usage — and, for a limited set of accounts, secrets, datasets and a subset of incomplete model weights. Payment card numbers are not stored by Fireworks and were not affected, and the company says it has detected no impact to Inference, the Training SDK or Nexus, which are zero data retention by default, though the investigation is ongoing.

Fireworks has disabled self-serve dataset uploads and downloads while the investigation continues, brought in Palo Alto Networks Unit 42, and notified law enforcement. It links the activity to a financially motivated group engaged in LLMjacking — reselling unauthorized access to large language models.

2026-11-08

ANNOUNCED

Nvidia's board opens an independent investigation into reports CEO Jensen Huang was targeted by a foreign intelligence effort

Nvidia's board of directors said on Sunday, November 8, that it is opening an independent investigation into reports that chief executive Jensen Huang was the target of a foreign intelligence effort. A committee of independent directors, assisted by outside counsel, will oversee the review. The board said it has 'no reason to believe' that either national security or Nvidia secrets were compromised, and that Huang fully supports the review, has committed to cooperating, and will remain CEO while it runs.

The announcement followed New York Times reporting that U.S. intelligence agencies determined Alice Huang, a Chinese citizen and former executive at Singapore-based data-centre company Megaspeed, was gathering information for Beijing. According to the Times, Huang acknowledged to the board that he had been intimate with her, and apologised in a message to staff and family. Megaspeed has bought more than $4.8 billion worth of Nvidia chips over the past two years, the Times reported; the two Huangs are not related.

Nvidia's chips power much of the AI build-out and sit at the centre of U.S. export restrictions on China, which is why a report of this kind reaches the board and Congress: three U.S. House committee chairmen have asked Huang to attend a closed-door briefing on November 18. China's foreign ministry said on Monday that the spy story was unsubstantiated, Reuters reported. Nvidia gave no timetable for the investigation.

ANNOUNCED

FT/Reuters: Anthropic took over $3.4 billion in gas-generator orders for 1.7 GW of data-centre power

Anthropic has struck a deal worth about $3.4 billion to secure gas-fired power generators for its data centres, the Financial Times reported on Sunday, in a story picked up by Reuters on November 8. The company bought commitments for 1.7 gigawatts of heavy-duty generators in late September from OnePWR Solutions, a San Antonio, Texas-based equipment supplier.

Anthropic did not acquire OnePWR itself, and none of the supplier's employees transferred, according to the FT's source — but the deal included the engineering and construction capabilities needed to install the generators. Taking over existing orders, rather than joining the back of the queue, is a way to get power equipment sooner at a moment when electricity, not chips, is becoming the binding constraint on new AI capacity.

Neither company has confirmed the transaction: Reuters said it could not immediately verify the report, and both Anthropic and OnePWR declined to respond to requests for comment. The figures — $3.4 billion, 1.7 GW, late September — are the FT's, reported through Reuters, and the environmental cost of gas-fired generation for AI campuses is already part of the argument around the deal.

2026-11-06

ANNOUNCED

Physical Superintelligence says an unreleased Anthropic model proved the Aaronson-Ambainis conjecture, open since 2008 — Lean checked the proof in 3 days

Physical Superintelligence (PSI) announced on November 6 that an internal research system running instances of a non-public Anthropic model produced a complete proof of the Aaronson-Ambainis conjecture, a quantum-computing question open since 2008. The system had a complete proof on the third day of its run on PSI's own infrastructure, and the Lean proof assistant accepted it; PSI then pointed the system at related problems and it improved the result twice before Lean accepted the final version.

PSI published both the human-readable paper and the Lean 4 formalization on GitHub. The formal development — hundreds of machine-checked theorems — contains no unproven placeholders or added axioms according to the repository's own audit, and its declarations were replayed through Lean's kernel and lean4lean, an independent type checker. David Turturean of PSI built and ran the research system, framed the problem and takes responsibility for the paper's correctness.

The caveats are explicit in PSI's own material. The paper is a preprint with no peer review yet, the prose paper is not itself machine-checked, and human experts still have to confirm that the formalized statement matches the conjecture Aaronson and Ambainis actually posed. PSI has not identified which Anthropic model it ran, and Anthropic has not commented.

ANNOUNCED

OpenAI partners with Hacktron to hunt for vulnerabilities in its own systems, with wider Daybreak Red access

OpenAI announced on November 6 that it is partnering with security firm Hacktron to look for vulnerabilities in OpenAI's own systems. Engineers from both companies will jointly pick high-priority targets, with a particular focus on the seams where OpenAI's systems connect to third-party services.

As part of the deal, Hacktron gets broader access to OpenAI's frontier models through Daybreak Red, the restricted program that also gates GPT-6 Cyber. Hacktron builds AI agents for code review, penetration testing and red-teaming, and OpenAI first encountered the firm through its Security Bug Bounty program, where its researchers reported flaws in how applications, cloud infrastructure and identity systems interact.

OpenAI placed the partnership alongside two other steps the same week: a security sprint with Cathedral's researchers and raised bug-bounty rewards. No findings, counts or scope list were disclosed, so for now this is a statement about how OpenAI intends to test itself — with an outside offensive team aimed at its integrations — rather than a result.

ANNOUNCED

Bloomberg/Reuters: OpenAI banks the first $11 billion of its raise at a $1.25 trillion valuation — well short of the $30 billion it first sought

OpenAI has secured the first $11 billion of its current funding round at a $1.25 trillion valuation, Bloomberg reported on November 6, with the funds transferred to the company by Friday. The round has not closed and could still reach about $20 billion in the coming weeks — the target OpenAI cut to in late October — but that is well short of the $30 billion bridge round first reported on October 22. Coatue Management and T. Rowe Price were among the participants in the first close, with SoftBank and MGX also reported in the round.

The smaller, slower raise is the clearest price signal yet that investor appetite for AI paper has cooled since the midterm sell-off. OpenAI first sought at least $30 billion at a valuation around $1.4 trillion; it is now filling the round partly from family offices and high-net-worth individuals, according to the Wall Street Journal's earlier reporting.

The funding lands alongside President Greg Brockman's comments at the WSJ Tech conference on November 4 that OpenAI could be IPO-ready in 2027, that enterprise revenue has overtaken consumer sales, and that OpenAI has formed a strategic partnership with JPMorganChase that includes access to its Daybreak cybersecurity models. OpenAI has not publicly confirmed the round's details. The Wire's earlier rumor entries on the $30 billion bridge and the cut to $20 billion stay unresolved until the round actually closes.