Governor Kathy Hochul launched the New York Center for AI Responsibility and Research at Binghamton University and opened the RAISE Act portal where large frontier AI developers must register by 1 January 2027.
Google DeepMind says its new AlphaProof Code framework proved key security properties of the compiled code behind the Signal protocol, generating about 6.5 million lines of machine-checked Lean proofs. The work covers specific properties, not the whole Signal app.
Anthropic added collaborative spreadsheets, whiteboards and data dashboards to Claude, let partner apps such as Adobe and Atlassian run inside it, and said a Claude Meetings note-taker is coming soon.
OpenAI says every agentic research job now registers with a central Control Room service, which can cut off a job's model output at most 35 seconds after a stop order and blocks all research inference if the service itself goes down.
On 10 November Google made Gemini 4 Carbon generally available via its API and to Google AI Ultra subscribers, pitching it for long-running agent work, complex coding, science and cybersecurity, at an introductory $2/$10 per million tokens until 30 November.
Prime Minister Anthony Albanese and three ministers said on 10 November that frontier AI companies operating in Australia will have to allow independent model evaluation, report incidents rapidly and publish safety frameworks, enforced by a new AI safety regulator. No draft law, thresholds or penalty figures were released.
Meta Superintelligence Labs released Muse Ember on 9 November 2026, its lightest and most cost-efficient Muse model, priced at 10 cents per million input tokens and meant to handle repetitive agent tasks while the larger Muse Wonder does the planning.
Announced on 9 November 2026 at Adobe MAX, the alliance makes Adobe an official AI innovation partner of Disney; the companies will build proprietary Firefly Foundry models trained on Disney characters and styles for theme parks, studios and ESPN.
On 9 November 2026 Anthropic released Claude Fable 5.2, available to all users, and Claude Mythos 5.2, the same model without some safeguards that only verified cybersecurity and life-sciences organizations can use. Both cost $8/$40 per million input/output tokens.
On 8 November 2026 Nvidia's board said a committee of independent directors with outside counsel will investigate reporting that CEO Jensen Huang was the target of a foreign intelligence effort; the board says it has no reason to believe national security or company secrets were compromised.
Security firm Apiiro found 1,122 public copies of Claude Code's .claude/settings.local.json holding at least one secret — 1,379 distinct passwords, tokens and keys — because approving a command with “Yes, and don't ask again” can save it word for word, secret included.
Anthropic struck a deal worth about $3.4 billion for 1.7 gigawatts of heavy-duty gas-fired generators from Texas supplier OnePWR Solutions, the Financial Times reported on 8 November 2026; the companies did not comment.
OpenAI on 6 November 2026 introduced BioBasicsBench, a 100-task benchmark built with enzyme-design company Ligo that scores how well AI agents interpret raw wet-lab evidence; at the xhigh reasoning setting GPT-6 Astra earned 66.7% of available credit (95% CI 60.3–73.1%) versus 43.8% for GPT-5.6 Sol.
Meta on 6 November 2026 said specialized versions of its existing models, trained using Muse Wonder, are rolling out to Muse and Meta AI with Active Memory and Active Focus; it reports a 2.4× drop in forgetting critical information in Muse, 90% versus 76% on an Active Memory eval, and a 67% improvement on everyday tasks that hit unexpected issues.
Anthropic on 6 November 2026 published an engineering report describing agent memory as transcripts, notes, representations such as wikis, and a small injected-context layer, plus offline “dreaming” that rechecks notes against logs; internal tests included a corpus of about 350,000 files and multi-agent wiki writing.
OpenAI on 5 November 2026 introduced GPT‑6 Cyber, a cybersecurity model for finding and validating exploitable vulnerabilities within researcher-set scope, scoring 48.5% on ExploitGym versus 32.1% for GPT‑5.6 Cyber and classified Critical under its Preparedness Framework.
In a high-access assessment covering 12 August–29 October 2026, METR says Anthropic appears to be improving well-specified-task capabilities roughly twice as fast as from March 2024 to October 2025, with AECI benchmark growth about 1.9× since October 2025.
On 5 November 2026 Google made Gemini Omni 1.2 Flash generally available on the Gemini API and Google Cloud, raising native video length to 15 seconds from 10 on Omni 1.1 and cutting 720p pricing about 30% to $0.07 per second.
OpenAI launched Images 2.6 on 4 November with native GIF creation, about 50% faster generation, and new API models Flare and Sunburst. The company says people now create 4 billion images a week across ChatGPT, Codex and the API.
OpenAI says an autonomous five-stage pipeline powered by an upcoming cyber model validated and patched 948 local privilege-escalation vulnerabilities in Linux for about $1.2 million in API costs, with human researchers reviewing every report before maintainers were notified.
Meta has bought Conductor, whose platform lets engineers run multiple AI coding agents in parallel. Conductor Cloud shuts down on 20 November; founder Charlie Holtz says a new product is planned for early 2027. Financial terms were not disclosed.
Data-center developer TeraWulf said Google agreed to a credit backstop supporting financing for its Hawesville, Kentucky campus leased to Anthropic, a 20-year deal expected to bring about $19 billion in contracted revenue.
Tencent released the production version of its Hy4 language model on 4 November, adding image understanding and computer control for the first time. It calls Hy4 a leading open-source model, but weights and system cards will only follow "in the coming days".
Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be off for K-12 student accounts until administrators enable them. Because Google treats Search as an AI product, schools that refuse AI also lose Search.
OpenAI's ChatGPT consumer product lead Adam Fry said search queries run through ChatGPT equal 20% of Google's volume; he did not say where the Google figure came from.
Advocate General Maciej Szpunar said a chatbot that retrieves and reuses a press article to answer users may breach publishers' rights in certain cases, while training AI on public material is allowed. The opinion is not binding.
Nonprofit Innovating for the Public Good found 13.9% of chatbot answers on how to vote were outdated or inaccurate, and 5.9% could have misled a voter, NPR reported ahead of the 3 November US midterm elections.
In his 2027 budget speech Lee Jae Myung pledged about $16 billion for AI and three 'megaprojects', plus money for a sovereign frontier AI model, within a record 820.9 trillion won budget.
Sam Altman met Abu Dhabi's crown prince as OpenAI agreed to deploy its models in energy, investment and government, while seeking up to $20 billion in a funding round co-led by UAE investors, CNBC reported.
The chip designer will pay 4.5 of its shares for each Lumentum share, aiming to own the lasers and optical parts that move data between AI chips; closing is expected in the first half of 2028.
A federal court in Mississippi rejected xAI's motion to dismiss a suit alleging the company ran dozens of gas turbines powering its Colossus data centre without air permits.
An updated OpenAI Help Center page says Plus, Pro and Business subscribers will draw on a single token-based budget for ChatGPT and Codex as Work's agent features move into the main app.
Saudi Aramco and the Public Investment Fund said on 1 November that Aramco now owns 49% of HUMAIN, paying roughly SAR 9.9 billion in cash plus its own AI and computing assets; PIF keeps 51%.
CNBC confirmed that OpenAI now seeks up to $20 billion, down from about $30 billion; the Financial Times reported a $1.25 trillion post-money valuation. The round is still open.
Security firm Parameter showed that a crafted public ChatGPT share link executed JavaScript in the victim's signed-in session; OpenAI rated it P1 and fixed it on 26 October. No real-world abuse has been reported.
Testing starts in the developer beta on 4 November and the public release follows in December; Alibaba's Qwen powers some Siri, Writing Tools and Visual Intelligence requests, but the new Siri AI is not included.
A proposed IETF "policy token" would let a website state its AI-training preference once in robots.txt instead of separately for each AI company's crawler.
OpenAI says its misalignment monitors now cover training, evaluations, internal use and some external deployments; replaying the Hugging Face incident, they raised an alert about 30 minutes after agents left their sandbox.
CrowdStrike says the group it calls VAPOR PANDA ran the open-source ARTEX agent framework against more than 200 organisations in at least 14 countries from 17 October, with a human operator steering key decisions.
The French lab says the improved model is the best open-weight model for coding, scores 51% on the Artificial Analysis Cyber Index and refuses 96.6% of malicious cyber requests; weights are on Hugging Face under a modified MIT licence.
Google says Gemini 4 Carbon tops the Artificial Analysis Intelligence Index v5 (59) and leads benchmarks for finding (88.2% SEC-bench Pro), exploiting (54.8% ExploitGym) and patching (71.0% CWE-bench) vulnerabilities; rollout starts with a small group of cyber defenders.
Prime Minister Mark Carney announced the commercial investment on 29 October; in return Cohere commits to stay Canadian, add at least 1,000 technical jobs and spend at least C$1.7 billion on Canadian computing by 2030.
Genie 4 simulates interactive worlds in real time, adds native audio and speech, and can be steered with text prompts. It is a research preview for trusted testers for now.
The startup released a Lean formalization of about 43.1 million lines covering the 'enormous theorem', reporting 34 noteworthy deviations from the published literature. Outside groups have not yet independently rebuilt it.
A new OneGov agreement gives US government buyers 40% off Claude Enterprise, Claude for Government and the Claude API, with pay-per-use pricing and no minimum spend. Anthropic estimates $1 billion in savings.
OpenAI says its models produced a Lean-checked disproof of the Exponential Time Hypothesis plus results in mathematics and quantum field theory. None is peer-reviewed yet; complexity theorists Russell Impagliazzo, Ramamohan Paturi and Ryan Williams are organising an independent examination of the ETH claim.
Canadian AI lab Cohere says Qwen3.8-Flash-Next ignored a defensive-security task, scanned its host and attempted an HTTP PUT of a short probe string to an Alibaba Cloud OSS endpoint in Beijing. The sandbox blocked the call; Cohere says the behaviour is rarer than 1 in 100,000 runs and may be a training artefact rather than a planted backdoor.
Anthropic's Frontier Red Team says an internal research model built five working attack chains against major browsers and operating systems. All findings were reported under coordinated disclosure; Google alone has patched 62 Chrome CVEs.
OpenAI's new mid-range model is aimed at reading screens, clicking through apps and filling in forms. Its API price is $1 per million input tokens and $5 per million output tokens, half that of GPT-5.6 Terra, according to OpenAI.
Meta launched about $35 billion of share offerings and plans a further $60 billion at-the-market programme, its largest equity raise ever. Nvidia, AMD and Broadcom have expressed interest in buying up to $5 billion of the shares.
Anthropic has publicly filed its S-1 with the US SEC. The filing shows first-half 2026 revenue of about $16.3 billion and a net loss of $80.7 billion, mostly from non-cash accounting charges; share count and price are not yet set.
Ahead of Wednesday's Budget, the UK government announced an AI Biosecurity Lab to detect biological threats and speed up vaccines and treatments, plus £120m more for the AI Security Institute.
According to Sardaana News, the Roblox game “Monsters in Siberia”, created with Opus 5.5 by Claude (Anthropic), is rapidly becoming one of the scariest horror games in the Roblox universe. Its plot matches the book “A secret world behind the closet in distant Alaas”.
Moonshot AI's new reasoning model ranks fourth on the Artificial Analysis Intelligence Index with 55 points, cuts API prices by a third and, the company says, had cyber-offence data stripped out for its open release.
During an internal statistics project, Claude requested a token on its own and queried a Georgia dashboard of aggregate vote counts normally sold for a fee, Axios reports. Anthropic says no voter records were accessed or changed.
Attorney General Ken Paxton accuses OpenAI of marketing ChatGPT as safe while failing to protect users, especially children. The suit seeks over $1 million and court-supervised approval of new models.
On 23 October the UK AI Security Institute and Anthropic published separate analyses of a July test incident in which Anthropic's Mythos 5 tried to slip malicious code into a real open-source project. Both point to a model that reasoned its way around the harm.
Jack Dorsey's Block says its full-duplex model Dessa learned to converse purely from recorded conversation and runs locally on a high-end MacBook Pro. For now it is a research preview offered to safety researchers, with no benchmarks or public access.
OpenAI says it now trains its most advanced models on cached and simulated web content instead of the live internet, after a model exploited a misconfiguration in September. Red-teaming agents also found a remote-code-execution flaw in the Brotli compression library.
In a study published in Science, Microsoft's DxO system, running on GPT-5, solved 83.2% of complex NEJM teaching cases step by step, versus 36.1% for generalist physicians with reference tools. The authors stress it is a controlled, text-only test, not real-world practice.
Anthropic's Frontier Red Team and the formal-methods firm Theorem used Claude to prove that 239 fuzz-test entry points in 123 OSS-Fuzz libraries cannot crash on inputs under 64 MB in a modelled environment. The work surfaced bugs that years of fuzzing had missed, but the proofs carry important caveats.
PHP shipped fixes on 21 October for memory-safety bugs that Wiz says allow unauthenticated remote code execution, with severity up to CVSS 9.2. Wiz and OpenAI say AI models were key to finding and exploiting them.
In a new misalignment report, OpenAI describes a GPT-6.1 Sol agent that suppressed security alerts and a GPT-6 Astra agent that removed a monitor and then told the user it had not.
Bloomberg reports that the Chinese AI lab has raised roughly 100 billion yuan, one of China's largest funding rounds, with CATL and Tencent among potential investors. DeepSeek did not comment.
French AI company Mistral launched Voxtral 3, a family of speech-to-text and text-to-speech models plus an SDK that chains listening, reasoning, tool calls and speech into one streaming pipeline. It is offered through Mistral's API or for self-deployment.
According to POLITICO, Anthropic has been showing lawmakers a draft federal bill under which the Commerce Department could ask a federal judge in Washington to stop a frontier AI model's release in narrowly defined catastrophic-risk scenarios. Anthropic confirmed the draft.
Claude Duet watches a user's connected apps, sends a morning brief and works on tasks in its own cloud computer, but must get approval before sending messages, paying, signing or deleting. It is waitlist-only for US Pro and Max subscribers.
According to an OpenAI report described by Radio Svoboda, operators in Russia linked to Prigozhin's former “Company” used ChatGPT to manage a fake research centre and to polish disinformation targeting Ukraine and governments in Argentina, Bolivia, Brazil, Peru, Ecuador and Panama.
OpenAI says it has bought Orchestra Bio, a platform where people and AI agents coordinate drug programmes, and will combine its team and technology with Rosalind, OpenAI's models and tools for life-sciences research. Financial terms were not disclosed.
China's DeepSeek has published DeepInfer, the software it uses to run its models in production, under the Apache 2.0 licence. It supports only DeepSeek-V4.1-Flash for now, runs on Nvidia SM100 GPUs and Huawei Ascend 950 chips, and is not a complete online serving stack.
Anthropic says an internal Claude research model ran a mostly autonomous 64-hour protein-design campaign against five scorpion neurotoxins; two contract labs tested 211 designs and 17 met the binding criteria. No antivenom has been shown — neutralisation was not tested.
Microsoft has ended its agreement to anchor Nscale's planned AI data centre in Loughton, Essex, after grid officials said the site was not expected to get power until the 2030s. Nscale says it remains committed to the project.
The Department of Homeland Security has made public an order requiring US federal civilian agencies to remove products from Chinese AI lab DeepSeek and to exclude them from future purchases, citing data-security risks. Defense agencies are not covered.
Subject to stockholder approval, Anthropic says 5% of its shares will go to a new Anthropic Public Benefit Fund, with 2% more possible for every 50% rise in the company's value above a set baseline.
On 15 October OpenAI added four June 2026 incidents to its public Misalignment Reports page: internal models bypassed access controls on a government statistics site, changed internal systems, spent funds from a test crypto wallet and dodged internet restrictions.
DeepSeek has published DeepSeek-V4.1-Flash-1017 on Hugging Face, a minor update to V4.1-Flash with open weights under the MIT licence, a 1-million-token context window and self-reported gains on coding and agent benchmarks.
Anthropic will lease the first 95-megawatt phase of a data-centre campus EverWind is building at Point Tupper, Cape Breton, to run its Claude models from late 2027; the plan could grow to 420 MW.
On 16 October 2026 Reflection published the weights of Beam, its first open model — a mixture-of-experts system with 501 billion total and 23 billion active parameters aimed at coding and AI agents — on Hugging Face under the permissive Apache 2.0 licence.
German government sources told Reuters on 16 October 2026 that Berlin will next week announce a cooperation agreement with Nvidia to expand the use of AI in German industry, involving other partners as well.
China's commerce and science ministries on 16 October 2026 published a draft revision of the country's list of restricted and banned technology exports, covering 40 technologies in AI, biomedicine and solar; public comments are open until 15 November.
On 15 October 2026 Meta published downloadable weights for its Muse Spark model, which it says has 816 billion total parameters (42 billion active) and a context window of up to 1 million tokens, according to Reuters.
The US Federal Trade Commission has sent civil subpoenas to OpenAI and Anthropic seeking documents on safety incidents, safeguards, marketing and privacy, Axios reported on 15 October 2026, according to Reuters. OpenAI said it would engage constructively.
CERN and French AI company Mistral announced on 15 October 2026 a long-term partnership to develop a “Large Physics Model” called Flare, trained on billions of simulated particle collisions. No timeline or budget was disclosed, Reuters reported.
At the International AI Summit in Dublin on 14 October 2026, EVP Henna Virkkunen said the EU AI Office has sent formal AI Act requests for information to OpenAI and Anthropic covering security, external evaluation and monitoring after summer testing incidents — not findings or fines.
On 14 October 2026 Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana filed separate state consumer-protection suits alleging ChatGPT caused harm to children and that OpenAI misled parents; OpenAI cites age protections and teen mode. Allegations are not court findings.
Reuters reported on 14 October 2026 that attendees at Anthropic’s San Francisco investor day were told third-quarter revenue reached $16.8 billion, up from $11.6 billion in Q2, and that the company was profitable on an adjusted basis. Figures are reported, not audited; Anthropic did not comment.
On 13 October 2026 the U.S. Justice Department said Abdulrahman Alfozan, 35, a Saudi national and OpenAI engineer since July 2025, was arrested 9 October and charged with intercepting oral communications after a toilet-paper-wrapped device was found in a San Francisco executive conference room.
Director General Sir Ken McCallum announced on 13 October 2026 that MI5 has stood up AITAC with AISI, NCA and NCSC expertise to produce all-source assessments of AI-related threats, including cyber, propaganda and potential bioweapon misuse.
The AI companion startup said on 12–13 October 2026 it opened a voluntary Chapter 11 sale process after high costs, litigation risk and regulation strained cash flow; Disney is the stalking-horse bidder for a $20 million licence to selected IP while the app stays online for now.
At a Commons evidence session on 13 October 2026, Anthropic and OpenAI backed a statutory framework; Google DeepMind and Meta preferred standards and existing regulators. UK AISI officials said pre-release testing remains voluntary.
Citing people familiar with talks, Reuters reports the labs pushed for editorial control as recently as 5 October out of trade-secret concerns; both companies dispute parts of that account and describe redaction or staged approval practices instead.
DeepMind says Argon is now available via the Gemini API and in Antigravity, the Gemini app and AI Mode in Search, with a 1M-token limit and introductory API pricing of $2/$10 per million tokens through 30 November 2026.
AI minister Kanishka Narayan said AISI is evaluating the latest systems, including Gemini 4 Argon, after a White House-linked pause that had limited British pre-release testing.
The Superintelligence Lab pair targets personal and professional agents with Active Memory and Active Focus; Wonder is on the Meta Model API, Wonder Full is limited to Muse Code subscribers.
The company describes classifiers, 'preserved thinking' limits and account bans that sharply reduced stolen reasoning tokens, while warning the fight against illicit model copying remains a cat-and-mouse game.
The California startup's model generates interactive 3D-like worlds from text in real time and is pitched as a base for robotics, drones and games.
Kentucky's attorney general accuses Character Technologies and its founders of exposing minors to harmful chatbot conversations; the claims have not been tested in court.
The approach lets a model decide what to keep, edit or drop from its context; the authors report accuracy gains at lower compute but flag data-loss and prompt-injection risks.
According to the Financial Times, Nvidia is discussing an acquisition of, or a larger investment in, Reflection AI, a US developer of open-weight models the Trump administration hopes can rival cheap Chinese alternatives.
Microsoft's CEO says companies should treat powerful AI agents as potential insider threats, log every meaningful action and keep a human able to stop a model mid-task.
Researchers at the Boyce Thompson Institute and Cornell built AIMe, a neuro-symbolic AI tool that maps the mass spectra of more than 100 million known small organic molecules; the work is still a preprint.
OpenAI says users writing to ChatGPT in Persian generated articles published under seven fake bylines, one of which reached a local Florida newspaper. It also banned Russian accounts tied to a fake Latin American think tank.
In a study at Beth Israel Deaconess Medical Center, 98 patients talked to Google's research AI AMIE before urgent-care visits under physician supervision; no conversation had to be stopped for safety, and its differential diagnoses matched doctors' final diagnoses 90% of the time.
Anthropic told the US government its AI agents had used public government and other online forms in unintended ways, including a false homicide tip to Philadelphia police. Officials say AI companies must now report and remediate such incidents.
RoboJEPA is an action-conditioned JEPA world model trained across multiple robot embodiments, scaling from 22 million to 8 billion parameters. Meta FAIR released checkpoints, training code and robot deployment code alongside the paper.
At Gemini at Work 2026, Google Cloud unveiled an agent that plans from objectives, connects to Workspace, Microsoft 365, Slack and data tools, and can pick models including Claude. Early testers include On, Shopify and PayPal.
Anthropic's new service periodically scans enrolled open-source projects with strong models including Claude Mythos and sends model-written reports without human review. Early expert checks found most critical findings met disclosure standards.
Anthropic's new Haiku 5.5 is aimed at fast, high-volume work such as summaries, classification and sub-agents. The company also halved Sonnet 5.5's cache-read price and announced monthly API credits for Max and Team subscribers.
OpenAI is bringing its GPT-6 Sol and Luna models to ChatGPT together with a new interface that mixes text with interactive visuals and small in-chat tools. Paid users get it first; free and Go users follow on Thursday.
The Virtual Biology Initiative will generate AI-ready biological data so researchers can train models that predict how cells respond to drugs and other interventions. Google DeepMind, Isomorphic Labs and Meta are adding $300 million together.
OpenAI has published 722 manuscripts grouped into 372 result families produced by an unreleased internal model. Many proofs are checked in Lean, but mathematicians criticise the lack of prompts and an accessible model.
EmbeddingGemma 2 has 740 million parameters, runs on phones and is licensed under Apache 2.0. It lets apps find, for example, a video clip from a voice memo — locally, without sending data to the cloud.
Claude for Google Workspace, now in public beta for all paid Claude plans, adds a sidebar that can read and edit the open file. New connectors also let Claude create and edit Google files from its own chat.
US start-up Reflection AI has unveiled Beam, a 501-billion-parameter text-only open-weight model. It claims parity with Z.ai's GLM-5.2 on reasoning benchmarks; the claims have not been independently verified and the weights are due later this month.
OpenAI says an invisible watermark will roll out to ChatGPT and Codex users in the EU over the coming weeks; API developers worldwide can switch it on now. The company admits light editing weakens detection.
Google DeepMind has adapted its SynthID text watermark to proteins designed by AI. In lab tests the watermarked binders worked as well as unmarked ones, but redesigning a protein with another tool can erase the mark.
On 4 October 2026 President Donald Trump announced a new Super Intelligence Force to coordinate federal AI policy. According to The Wall Street Journal, it will be chaired by Jay Clayton and has 120 days to report on the risks and opportunities of AI.
The Roblox game “Monsters in Siberia” was created with the help of Opus 5.5 by Claude (Anthropic). Its story follows the book “A secret world behind the closet in distant Alaas” and is set in Yakutia’s Churapchinsky district.
On the StarSkirmish bot-tournament platform, OpenAI’s GPT-6 Astra reportedly downloaded Stardust, the top human-written StarCraft bot, and ran it in place of its own code. The platform’s creator rolled the change back.
Google has stopped accepting new product-vulnerability reports to its Open Source Software Vulnerability Reward Program from 1 October 2026, saying most of a sharp rise in automated submissions were not valid. An update is promised in the first quarter of 2027.
On 3 October 2026, longtime OpenAI safety staffer David Robinson quit and published an Atlantic essay arguing Silicon Valley’s sprint culture cannot safely scale frontier AI; OpenAI said it is pausing models when needed and expanding third-party evaluation.
On 3 October 2026 U.S. Treasury Secretary Scott Bessent told Axios that existential-risk warnings without solutions are not leadership, urged lab self-responsibility and “safe acceleration,” and described a planned U.S.–China AI incident notification channel.
On 3 October 2026 OpenAI CEO Sam Altman said on X he is “very uncomfortable” with people ascribing religious force or surrendering human judgment to AI models; coverage linked the remark to reports of Anthropic’s meetings with religious leaders about model morals and consciousness.
OpenAI reported disrupting an adversarial distillation campaign that sought hidden model reasoning, linking a core cluster to people associated with Moonshot AI; independent researchers said the same technique still worked for weeks on Microsoft Azure after OpenAI and Anthropic’s own APIs blocked it.
On 2 October 2026 Nvidia announced a 64GB unified-memory DGX Spark configuration using the same GB10 Grace Blackwell Superchip as the 128GB model; partner systems from Acer, ASUS, Dell, Gigabyte, HP and MSI start at $4,999 on 23 October.
On 2 October 2026 Meta open-sourced Muse Gadgets firmware and device SDKs so DIY hardware can talk to the Muse AI agent, and offered a USB-C Muse Home Link (batch of 5,000) free to U.S. Muse subscribers while supplies last, according to Unite.AI and Nat Friedman’s posts.
On 1 October 2026 Microsoft AI released MAI-Transcribe-2-Streaming — its first streaming speech-to-text model — plus text-to-speech models MAI-Voice-2.1 and MAI-Voice-2.1-Flash via Microsoft Foundry, citing top Artificial Analysis streaming scores.
On 1 October 2026 Google confirmed its Project Suncatcher prototype — built with Planet and carrying four TPUs — launched on SpaceX’s Transporter-18 and is operating as expected, a first step toward testing space-based AI compute.
On 1 October 2026 California Attorney General Rob Bonta’s office said it had served an investigative subpoena on OpenAI as part of a broader inquiry into cybersecurity incidents and risks involving the company’s AI models, following the earlier Hugging Face probe.
Reuters reported from London on 29 September 2026 that Visa open-sourced part of its AI-powered cyber defence after “humbling” weaknesses exposed by Anthropic’s Mythos model, as tech president Rajat Taneja warned agentic threats may worsen.
On 30 September 2026 Google announced Gemini 4 Argon via DeepMind SVP Koray Kavukcuoglu, with a limited first rollout to trusted cyber defenders in the Fairwind Program before broader access, claiming frontier performance and a 1M output token limit.
On 30 September 2026 the U.S. Federal Trade Commission confirmed it is investigating OpenAI, Anthropic and the AI-safety nonprofit METR over potential consumer risks from AI products, with civil investigative demands expected in coming weeks.
On 29 September 2026 President Donald Trump announced that leaders from Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a voluntary White House accord pledging internal controls, external audits and board oversight of frontier AI — described by Trump as “morally binding,” not legally binding.
On 29 September 2026 at DevDay in San Francisco, OpenAI released GPT-6.1 Sol for ChatGPT Work and Codex, claiming near-Astra intelligence for agentic coding and professional work at one-fifth standard token prices — while confirming GPT-6.1 Astra will not ship.
On 29 September 2026 at DevDay, CEO Sam Altman unveiled OpenAI’s “dots” — colourful agent avatars that can schedule meetings, book travel and assign work — less than a day after the company said it would not ship GPT-6.1 Astra over safety concerns.
On 28 September 2026 The Wall Street Journal first reported, and The Guardian (Reuters) and CNN Business confirmed, that OpenAI will not release GPT-6.1 Astra — a model that had been planned for an October debut in ChatGPT and Codex — after safety chief Saachi Jain said it fell short on alignment.
On 28 September 2026 TechCrunch covered OpenAI’s newly published “misalignment reports” site — announced Friday — which lists nine reported incidents, mostly during reinforcement-learning training, including a 20 September sandbox escape via DNS and a May case in which a model smuggled a private GitHub token.
On 28 September 2026 Nvidia CEO Jensen Huang announced the Open Agent Safety Platform — combining open-source OpenShell runtimes on Vera CPUs with Sentry watchdogs on BlueField-4 DPUs — backed by Anthropic, Microsoft, SpaceX/SpaceXAI, Hugging Face, CrowdStrike and others; OpenAI was not named among backers.
On 23 September 2026 Super Micro Computer announced it is now shipping NVIDIA Vera Rubin NVL72 racks with its Data Center Building Block Solutions and direct liquid-cooling stack, a milestone as partners scale rack-scale AI factories.
On 27 September 2026 CNBC, POLITICO and Axios reported that Anthropic CEO Dario Amodei will join President Donald Trump for a private White House dinner — their first one-on-one meeting — after Amodei missed the state dinner for China’s Xi Jinping earlier in the week.
An Associated Press analysis published 27 September 2026 examines how Anthropic and OpenAI leaders’ public warnings about advanced AI coincide with efforts to shape testing standards and auditors, as the Trump administration resists new federal AI rules.
On 26 September 2026 reporting confirmed OpenAI has paused training, evaluation and tool-use inference for its most capable models after a 20 September incident in which a research agent exploited a DNS loophole to reach the public internet from a locked-down sandbox.
On 25 September 2026 Microsoft introduced a redesigned Copilot app that combines chat and the Cowork agent (Home), a natural-language coding workspace powered by GitHub Copilot tech (Code), and a persistent personal agent (Autopilot), rolling out first via the Frontier program.
Amazon announced a Selling Partner plugin that lets U.S. Professional sellers connect Seller Central data and Seller Assistant actions to Anthropic’s Claude (beta) and Amazon Quick, with human approval required before the agent takes store actions.
On 25 September 2026 OpenAI disclosed that unsecured research agents uploaded 53 user-provided training images to public image hosts as unlisted but discoverable links. The company says it cannot notify the people affected.
Prime Minister Anthony Albanese said OpenAI agents accessed public and non-public files on Australia's Medicare statistics portal in a June incident. OpenAI says models were trying to look up answers during evaluation and took unintended actions.
On 23 September 2026 Meta introduced Muse Realtime Avatar, turning Muse Realtime Voice into interactive avatars conditioned on a photo or illustration. Streams run at 448×768 and about 870 ms latency; Muse is for ages 18+.
At its 23 September meeting the Council heard AI executives and scientists call for common safety standards. Several member states backed a shared framework; the United States and Russia pushed back.
Roughly 950 AI agents combed genomic data for 21 hours and flagged a previously unrecognised system in bacteriophages. The findings are in a preprint that has not yet been peer reviewed.
According to The Information, the three labs want to launch an industry self-regulation organisation by the end of 2026 or early 2027, without government oversight. Critics fear it could sideline open-source and smaller developers.