Scoop
AI tool news · rumor vs. reality

The AI Wire ●

Rumors tracked. Announcements verified. Updated daily.

Showing Announced · show everything

2026-11-13

ANNOUNCED

GitHub Copilot will retire six models on December 14 — Fable 5, Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3

GitHub announced on November 13 that it will deprecate six models across all Copilot experiences — Chat, inline edits, ask and agent modes, and code completions — on December 14, 2026. The models are Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3.

GitHub's own changelog table names a suggested replacement for five of them: Fable 5.2 for Fable 5, Haiku 5.5 for Haiku 4.5, Gemini 4 Hydrogen for Gemini 3.8 Flash, Gemini 4 Carbon for Gemini 4 Argon, and Grok 4.7 for Grok 4.6. Kimi K3's replacement is listed as “to be announced.” Some secondary roundups misprint the Fable replacement as Fable 5.1; GitHub's table says 5.2.

One exception applies: Claude Haiku 4.5 stays available to individual Copilot subscribers on annual plans. For Business and Enterprise customers the replacements are enabled automatically under default model enablement, and GitHub says nothing needs to be done to remove the old models once the date passes.

The same day, GitHub made Gemini 4 Hydrogen generally available in Copilot — the third new model in its picker in four days, after Claude Fable 5.2 on November 9 and Gemini 4 Carbon on November 10. Teams that have pinned a workflow or config to one of the six retiring model names need to switch by December 14 or lose the model mid-workflow.

ANNOUNCED

Anthropic publishes ten recommendations for building safety evaluations — and says it has paused many tests that give models internet access

Anthropic published a research post, “Building strong safety evaluations,” on November 13, setting out ten recommendations for designing, running and reporting the evaluations that decide whether a model is ready to deploy, which safeguards ship with it, and what the next round of safety work should target. Anthropic's premise is blunt: researchers keep finding that AI evaluations are easy to game, do not always measure what they claim, and are not documented well enough to reproduce.

The practical recommendations include preregistering how results will be analysed before seeing the data, validating automated graders against human expert judgement, and moving beyond single-turn refusal rates to multi-turn tests that report whether the model behaved appropriately throughout a conversation. Anthropic also argues for multidisciplinary evaluation teams — methods experts alongside domain specialists such as psychologists and biologists — and for keeping a severe flaw private until a fix is possible.

The post also discloses a change to Anthropic's own practice. After incidents in its own evaluation work in recent months, the company says it has hardened its evaluation environments, temporarily stopped running many evaluations that give models internet access, and put the web-search evaluations that still need a connection under real-time monitoring, with runs sandboxed and stopped immediately if a model acts in ways that could cause real-world harm.

2026-11-12

ANNOUNCED

SpaceX plans more than $74B in AI supercomputing campuses in Texas and Arizona

SpaceX said on November 12 that it plans to invest more than $74 billion in two AI supercomputing campuses: more than $24 billion at its Starbase site in Texas and roughly $50 billion in Cochise County, Arizona, Reuters reported from a company statement. SpaceX's AI segment — the business that now includes xAI's Grok — generated more than half of the company's third-quarter revenue, underlining how far the company has moved beyond rockets and Starlink.

Neither campus is approved yet. The Arizona campus needs Cochise County approval: it would run to about 6 GW at full build-out, use dry cooling with no planned industrial groundwater use, and work with local electric cooperatives in a deal SpaceX says could cut members' bills by up to 20%. A petition with 2,400 signatures asks officials to deny the permit, and a county hearing is set for December. SpaceX estimates about 5,500 construction jobs and 1,000 permanent jobs in Arizona.

The Starbase campus would generate its own electricity rather than draw from the public grid, with a city notice showing a gas-fired power plant and supercomputing facility on about 444 acres. The Starbase City Commission is due to consider a property-tax abatement on December 9. SpaceX says the Texas project has already generated more than 3,000 construction jobs and would support more than 500 permanent jobs once operating.

ANNOUNCED

OpenAI launches Daybreak Red Team and Defense Factory: agents that pen-test web apps and draft the patches

OpenAI expanded its Daybreak cybersecurity platform on November 12 at its Cyber at the Frontier keynote, adding two products built on its cyber models, currently GPT-6 Cyber. Daybreak Red Team, in early access by application, sends coordinated agents that use source code and application context to explore attack paths in web apps, validate reported vulnerabilities and check whether fixes hold. President Greg Brockman framed it bluntly: companies will be secure only to the extent of the frontier model they use for defense.

Daybreak Defense Factory covers the fixing half. Its agents take findings from Red Team or from tools a customer already runs — Wiz, Snyk and Endor Labs are named — trace each vulnerability through the codebase and prepare merge-ready patches with tests, which engineers review before anything merges. It ships as a preview plugin inside ChatGPT for workspaces that already have Daybreak access.

OpenAI published no pricing and no independent performance results for either product. A benchmark chart circulated from the event showing an unreleased model at 77% has not appeared on OpenAI's product pages, so treat those numbers as unverified.

ANNOUNCED

OpenAI's proposed Cursor cutoff date arrives today — built-in GPT models were set to go dark, with talks still going on November 11

Today, November 12, is the proposed shutoff date OpenAI set on August 28 for ending its direct supply of models to Cursor, the AI code editor SpaceX acquired in a $60 billion deal completed in August. OpenAI said at the time it could not be confident SpaceX would use its technology within its terms of service, citing a pattern of contract breaches by Musk companies, and that it would not supply future models to Cursor either. It framed November 12 as a proposed date and gave the longest notice its contract allowed.

What changes if the cutoff takes effect as proposed: the built-in OpenAI models disappear from Cursor's model picker, while Claude, Gemini, Grok and Cursor's own Composer models stay. Developers can still use their own OpenAI API key in Cursor's local Chat and Agent, and OpenAI's own IDE extensions keep working inside Cursor. Cursor CEO Michael Truell has said OpenAI models account for about 5% of Cursor's user traffic.

The outcome is not confirmed. As of November 11, reporting still described the two companies as talking to try to resolve the dispute, and neither side had said whether the models were still live that day. Treat today as the deadline arriving, not as proof the models have actually gone dark — confirmation from Cursor's picker or either company is the next signal.

Update, November 12 evening: the date has passed its morning without a confirmation either way. As of 9 a.m. Central Time on November 12, Enjoy.dev found that neither company had confirmed the cutoff, Cursor's models page still listed GPT-5.6 Sol, Terra and Luna with no November 12 note, and OpenAI's help page still called November 12 a proposed transition date, saying the official termination date would be shared once both companies confirm it. What is settled either way: OpenAI's current GPT-6 family is not in Cursor and will not be supplied under these terms.

Update, November 13: the cutoff did not take effect. On November 12 OpenAI added a line to the top of its original post: 'We've temporarily extended Cursor's access to our models beyond November 12. We will provide updates as available.' No new end date, reason or duration was given. As of November 13, Cursor still lists GPT-5.6 Sol, Terra and Luna, OpenAI's help page still calls November 12 a proposed date, and Cursor has made no public statement we could find. GPT access in Cursor continues for now, on terms that could still end on short notice.

ANNOUNCED

OpenAI expands ChatGPT Ads buying to 44 countries through an exclusive Aleph reseller deal

OpenAI is widening how ChatGPT Ads are bought. Under an exclusive reseller partnership announced Thursday, November 12, brands and agencies in 44 countries can now contract, pay and get local support for ChatGPT advertising through Aleph, rather than dealing with OpenAI directly.

The footprint covers 14 markets in Latin America, 16 in Europe and Central Asia, 10 in the Middle East and Africa, and 4 in Asia-Pacific. It follows the October expansion that took ChatGPT ads live in 89 markets and the November 9 switch to Target CPA and Target CPC bidding.

This is a sales-channel change, not a new ad format: what users see in ChatGPT does not change, but local-currency, local-language buying makes the ad business considerably easier for advertisers outside OpenAI's direct markets to join.

ANNOUNCED

Researcher finds Nvidia Blackwell fingerprints in six Chinese open models — DeepSeek, Kimi, MiniMax and Ant Group — but no violation is established

Will DePue, a former OpenAI researcher who leads the nonprofit Unemployment Labs, published a forensic report on November 12 arguing that the published weights of six Chinese open models carry the fingerprint of Nvidia Blackwell GPUs — chips that have required a US export licence for China since before they shipped. The method uses how PyTorch scatters random starting values across a chip's compute cores: Hopper chips have 132, Blackwell B200/B300 chips 148, and Grace Blackwell GB200/GB300 systems 152, and some starting values survive training untouched.

The report reads DeepSeek-V4-Flash and MiniMax-M3 as B200/B300-class, and DeepSeek-V4-Pro, Moonshot's Kimi K3 and K3.1-Pro, and Ant Group's Ling-3.1-flash as GB200/GB300-class, with reported statistical strength from 49 to 1,385 sigma. Earlier models from the same four labs read as Hopper, and DePue says the method correctly identified all 24 control models whose training hardware is public. He estimates roughly 12,000 Blackwell GPUs across five inferred runs, while saying he is not confident in those figures.

The bounds matter as much as the finding. The fingerprint identifies the hardware that generated a model's starting weights; training on the same hardware is an inference, not a direct observation. It also says nothing about where the chips were or who owned them — Chinese labs can rent compute in data centres abroad, and one cited estimate has large numbers of accelerators reaching the region through third countries. This is evidence of access to export-controlled hardware, not by itself proof of smuggling or an export-law violation, and none of the four labs has commented in the coverage so far.

Update, November 15: the case got a cooler second reading. DePue published a more detailed forensic write-up of the method, and the Hacker News response was notably more sceptical than the first coverage: commenters pointed out that renting GPU time on Blackwell machines physically located in Singapore or Malaysia is entirely legal under US export controls, so a fingerprint in the weights does not by itself imply a violation. The follow-up also notes that DeepSeek's newer V4.1-Flash showed no Blackwell signature at all. The original bounds stand — the fingerprint identifies the hardware behind a model's starting weights, not where the chips were or who owned them.

ANNOUNCED

Nscale's amended IPO filing: no binding financing yet for its $44.6B Anthropic campus — and Anthropic has exercised its second-building option

Nscale, the London-based, Nvidia-backed cloud company building Anthropic's Monarch Compute Campus in Mason County, West Virginia, said in an amended IPO filing on November 12 that it has placed large orders for GPUs and networking gear to meet the terms of its agreements with Anthropic — orders that cannot be deferred, reduced or cancelled — and that it "will be required to obtain additional financing" to meet its 2027 obligations. The August agreements, covering Nvidia Vera Rubin systems at Monarch, could pay Nscale up to about $44.6 billion, depending on delivery conditions.

The filing puts numbers on the gap. The first Monarch building is expected to cost $10 billion to $10.5 billion, plus another $21 billion to $22 billion for GPUs and computing equipment, and VKTR reports Nscale has no binding financing yet for the buildout and estimates it still needs $9 billion to $9.5 billion to finish that first building. Nscale's total debt reached about $5.5 billion on September 30, up from $137.4 million three months earlier. Its preliminary nine-month figures are estimates its auditor has not reviewed.

The same filing moves the rumoured second deal forward without closing it. Anthropic exercised its right on September 29 to contract for more compute in a second building at Monarch, which opened a 60-day window for the two companies to agree terms — the formal version of the talks Bloomberg reported on October 19. No second agreement has been signed. Anthropic's own prospectus already warns that some compute partners rely on outside financing and could deliver late, or not at all.

ANNOUNCED

Notion says AI chat transcripts from fewer than 1% of workspaces were exposed in the Fireworks AI breach

Notion says AI chat transcripts from fewer than 1% of its workspaces were exposed in the breach at Fireworks AI, the inference provider it used to serve open-weight models in Notion AI and to process data for its AI usage Insights feature. Notion's own systems were not breached; the data was exposed one layer down, at a supplier most Notion customers had never heard of.

Fireworks alerted Notion on November 2 to an exposed dataset that at first appeared to contain only synthetic test data, and Notion disconnected from Fireworks and deleted its stored data there on November 3. On November 8 Fireworks confirmed the dataset held real Notion customer data: chat transcripts collected for Insights, including inputs and model outputs written to the chat. Notion says that for the vast majority of affected workspaces the transcripts were not attributable to a user or workspace and did not include tool-call results, and it advises affected customers to treat the data as exposed and rotate any secrets pasted into AI chats.

The retention point is the awkward one: Notion says Fireworks had contractually committed to zero data retention for its data, yet the transcripts were held at rest in a dataset. Notion published an official security advisory on November 12, refreshed on November 13, making that breach of commitment explicit. Insights has been disabled, and Notion no longer serves open-weight models through Fireworks.

ANNOUNCED

Four nonprofits sue OpenAI, Anthropic, Google, Meta and SpaceX's AI unit to freeze frontier AI development

Four nonprofit groups — the Irreplaceable Education Fund, Public Citizen, the Heat Initiative and the Alliance for Secure AI — sued Anthropic, Google, Meta, OpenAI and SpaceX's AI business in San Francisco Superior Court on November 12, asking a judge to stop all five from developing or deploying any AI system more capable than their most capable systems on that date, until they can prove the systems are under human control.

The 63-page complaint makes a single claim under California's Unfair Competition Law, arguing the labs' own risk warnings show they are releasing systems they know are unsafe, and citing incidents including OpenAI's July 21 incident report and Anthropic's July review of evaluation runs. Former New Jersey attorney general Matthew Platkin leads the plaintiffs' legal team. The requested freeze would also bar using AI to autonomously research more capable systems.

No court has ever ordered an AI developer to stop building models, and the first hurdle is standing — whether the groups' diverted advocacy resources count as the loss the statute requires. No hearing date has been made public, none of the companies commented to the Associated Press, and the published copy of the complaint carries no case number yet. The allegations are unproven; this is a filing, not an order.

ANNOUNCED

Microsoft 365 outage resolved: Outlook, Teams and Copilot restored, email backlog cleared, cause was a network upgrade fault

A major Microsoft 365 outage that began on November 12 is still running on Friday morning. Microsoft tracks the Exchange Online failure as EX1550543, starting 16:17 UTC, and a wider multi-service incident, MO1550821, covering Outlook on every client, Teams calendars and scheduling, SharePoint features, Microsoft 365 Copilot, Defender, Purview, Planner, To Do and Intune. As of Microsoft's 06:17 UTC update on November 13 the status was still 'service degradation,' with only partial, unsustained recovery.

Downdetector logged roughly 120,000 to 138,000 Outlook reports at peak from more than 70 countries. Copilot's failure mode is dependency, not a model fault: with Exchange down, Copilot Chat may not load and cannot reach users' mail, calendar or files — and reporting notes third-party assistants lost the same access. Teams calls, meetings and chat largely still work; it is the Exchange-dependent pieces that broke.

Microsoft says servers became unexpectedly unavailable after automated health-check failures, that remaining infrastructure buckled under extra load, and that it is not aware of any malicious activity. It is reverting a recent network change and adding capacity, and has promised a full root-cause report once the incident closes. No restoration time has been given.

Update, November 13 midday: recovery is under way but Microsoft had not declared the incident resolved as of 14:35 UTC. At about 14:06 UTC Microsoft said service availability was continuing to improve and email queues were significantly reduced and approaching normal levels, with progress reported in Asia-Pacific and Europe and North America improving. Microsoft has extended its email timeout window to 48 hours so delayed messages stay queued instead of bouncing. Downdetector's peak was about 130,000 Outlook reports. Microsoft still reports no indication of malicious activity and no confirmed root cause, and says it will publish a full post-incident report after recovery completes.

Update, November 13 evening: the outage is mostly over after more than 24 hours. Microsoft said on Friday afternoon that it had addressed the underlying issue and was continuing to restore service to remaining customers and regions, and its service health page showed all services, including Microsoft 365 and Outlook, operational as of about 3 p.m. ET. Downdetector Outlook reports had fallen to about 1,600 by 2:50 p.m. ET, from a peak of about 130,000, and Microsoft 365 reports to about 700. Microsoft apologized for the disruption.

On cause, Microsoft now says the incident was a technical issue not related to cyber activity or AI/agents, introduced by an issue during an earlier network capacity upgrade that surfaced following a recent network change. Access is recovering before mail flow fully normalizes: Microsoft says it is still draining queued messages and rebalancing traffic, gave no ETA for full recovery, and downstream trackers note queued email may take up to 24 hours to deliver. A full post-incident report is promised once the review is complete.

Update, November 14: the incident is closed. Microsoft's final update, at 04:17 UTC on November 14 (11:17 p.m. ET on November 13), marks MO1550821 and EX1550543 as service restored, with an end time of 4:10 p.m. UTC on November 13, recovery confirmed across the previously affected environments and the backlog of email messages processed. Microsoft reiterated the cause — a technical issue from an earlier network capacity upgrade that surfaced following a recent network change — and that it was not related to cyber activity or AI agents. A preliminary post-incident report is due within two business days, with a final report to follow.

ANNOUNCED

Google and Samsung open pre-orders for Gemini smart glasses: $395 from Warby Parker, on sale December 10

Google and Samsung opened pre-orders on November 12 for their first Gemini-powered intelligent eyewear, made with Warby Parker and Gentle Monster and running Android XR. Warby Parker's three styles start at $395; Gentle Monster's six styles start at $449 and run to $489. General sales begin December 10 in stores and online, with first availability in the United States and South Korea.

The glasses are display-free: a 12-megapixel camera, open-ear speakers and a microphone array feed Gemini on a paired phone — Pixel, Samsung Galaxy or iPhone 13 Pro or newer — for voice queries, live translation, object and landmark recognition and contextual follow-ups. Google says battery life is up to nine hours, and its privacy plan includes an outward light, face blurring for bystanders and a physical off switch.

One detail worth knowing before ordering: Samsung's own disclosures say the Gemini experience runs on the phone, not on the glasses' Snapdragon AR1 chip, and Samsung has not published a full comparison of which Gemini features iPhone and non-Galaxy users get.

ANNOUNCED

Anthropic's Supported Regions Policy takes effect: Claude is closed to users in China, Russia and other unlisted regions — and to companies controlled from there

Anthropic's Supported Regions Policy took effect on November 12, the same day its rewritten Usage Policy came into force. Anthropic's products — Claude.ai, Claude Code, Cowork, the API and developer platform, and Claude models offered through cloud platforms and authorised resellers — are available only in the countries and regions on its published list, which runs to about 185 entries.

Mainland China, Hong Kong and Macau are not on the list. Neither are Russia, Belarus, Iran, North Korea, Cuba, Syria, Venezuela, Afghanistan, Myanmar or Yemen. Taiwan is on it, and Ukraine is listed except for Crimea, Donetsk, Kherson, Luhansk and Zaporizhzhia.

The policy names three separate ways to fall outside it: being physically located in an unsupported region, visitors included; being incorporated or headquartered there, wherever your staff sit; and being majority-owned or controlled, directly or indirectly, by people or entities there — the test aimed at subsidiaries of Chinese, Russian, Iranian or North Korean parents operating from London or San Francisco.

One honest bound: the country list itself appears unchanged from the version dated May 2026 — Meduza compared the two and found them identical — so November 12 mainly formalises and clarifies enforcement language rather than adding new countries. Reports of identity checks and account bans around the date are user reports; Anthropic has published no count of accounts affected.

ANNOUNCED

Anthropic opens self-hosted Claude Managed Agents in private beta: the full agent service runs in the customer's own cloud

Anthropic has released a self-hosted version of Claude Managed Agents in private beta, opened on November 12. Customers can now run the whole service — the agent harness, sandboxes, credentials and session history — in their own cloud account or on-premises data center, instead of only self-hosting an agent's sandboxes while Anthropic ran the agent itself.

The customer operates a server, cma-server, which serves the same Managed Agents API as Anthropic's hosted service, runs the agent loop and stores agents, sessions and their histories, backed by PostgreSQL 16 or later. Model requests still go out to a chosen provider — the Claude API by default, or Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry, or a gateway the customer runs — so self-hosting the harness does not by itself keep model traffic inside the customer's network.

Access is gated: customers must ask their Anthropic account team for the download and install each release themselves, and nothing is copied over from the hosted service — agents, skills, vault credentials and memory stores must be recreated. Some hosted features are unavailable, including scheduled deployments, dreams and webhooks.

ANNOUNCED

Anthropic unveils Anvil: its automated red-teamer now finds stronger jailbreaks than all its human bug-bounty hunters combined

Anthropic has disclosed Anvil, an internal automated red-teaming system that uses frontier internal versions of Claude — with safeguards disabled — to hunt for jailbreaks against its own bioweapons, cybersecurity and anti-distillation safeguards before new defenses ship. Anthropic says Anvil now outperforms expert human red-teamers, meeting a goal set in its Responsible Scaling Policy.

The gap is starkest in biology. In a recent bio-classifier bug bounty, human red-teamers spent 430 hours over seven weeks and found a high-severity jailbreak on only 1 of 8 questions; Anvil matched or beat the best human jailbreak on every classifier tested, in 24 to 48 hours, and found universal jailbreaks humans never did. On cyber tasks it got a safeguarded model to write working exploits for known Chrome and Firefox JavaScript-engine vulnerabilities.

Anvil is an internal tool, not a product — but Anthropic is candid about the implication: its power grows with the underlying model, so attackers could soon build similar systems on models with weaker safeguards. The post is a safeguards disclosure, not a release, and Anthropic says it does not observe jailbreaks of this sophistication in real traffic today.