DeepMind safety researcher Josh Engels joins METR, warning of severe AI harm
Josh Engels has left DeepMind's AGI safety team for METR, calling for slower development and independent investigation of AI behaviour.
AI News
AI news
The desk's complete reporting record, newest first.
- 94
- AI stories
- 12
- security flagged
- 05 Nov 2025
- archive starts
Amodei's AI slowdown call draws rival support and a global sell-off
Dario Amodei's call to slow frontier AI gains backing from Sam Altman and Elon Musk as chip shares fall and the US-China dispute sharpens.
Microsoft proposes human-control rules for its future AI models
Microsoft opens a six-week consultation on its Humanist AI Code of Conduct, proposing shutdown controls and safety rules for model development from 2027.
Why AI keeps picking 17: We tested randomness across 12 models
YFarmX tested 12 language models across number ranges, five languages and coin flips. Their answers reveal strong favourites and an unusually tidy idea of randomness.
Hermes joins Omarchy's desktop and terminal workflow
Omarchy's AI integration connects Hermes Desktop, the terminal command and theme settings around a shared runtime.
Xiaomi opens MiMo Desktop beta for editable work and computer control
MiMo Desktop's invite beta brings document creation, live previews and browser control. Xiaomi's access page currently excludes the UK and EU.
ChatGPT Images 2.5 adds sketch-led creation and more precise edits
Images 2.5 introduces Sketch, image comments and shareable prompts, with Flare and Sunburst bringing two options to the API.
OpenAI shares a proposed Navier-Stokes proof as researchers debate credit
OpenAI has released a forced Navier-Stokes blowup construction and Lean formalisation, alongside a dispute over research priority and private data.
DeepSeek plans at least 160,000 Huawei Ascend chips in Inner Mongolia
Bloomberg says DeepSeek wants at least 160,000 of Huawei's Ascend 950DT accelerators for a data centre at Ulanqab in Inner Mongolia. Huawei sells that chip for decode and training. DeepSeek's plan uses it for decode, and keeps Nvidia for the training runs.
HiddenLayer raised $100m, Wonderful $550m, and Bloomberg reports a $13bn Crusoe cloud deal
HiddenLayer raised $100m and Wonderful $550m, both announced on 2 September by the companies themselves. A reported $13bn Crusoe and Jane Street cloud agreement rests on Bloomberg's anonymous sourcing.
ARC Prize scored GPT-6 Astra at 62.7% and 99.9% on the same benchmark on the same day
OpenAI released GPT-6 Astra on 3 September. ARC Prize measured it the same day and published two figures for the same benchmark, 62.7% and 99.9%. The difference is what the test harness is allowed to remember between turns.
GenAI.mil added ChatGPT and Grok on the same day, both accredited at Impact Level 5
The Pentagon added OpenAI's ChatGPT Mil and xAI's Grok for Government to its GenAI.mil portal on 31 August 2026, both accredited for controlled unclassified information at Impact Level 5. Neither release names a single person.
Adobe values a year of free Firefly for 27 million people in Saudi Arabia at over $4bn, and publishes no method
Adobe said at LEAP 2026 on 31 August that it will give over 27 million Saudi citizens and residents 12 months of free Firefly and Express Premium, and valued the commitment at over $4 billion. The figure is Adobe's own valuation, and it appears in no filing.
Brussels has designated ChatGPT a very large online search engine, on OpenAI's own count of 159.1 million users
The European Commission designated ChatGPT a Very Large Online Search Engine on 31 August 2026, alongside Reddit and Roblox as platforms. The designation rests on a user number OpenAI reported itself, and the compliance clock runs from notification.
Hermes Agent now browses with a copy of your real logins and cookies
Hermes Agent shipped real-profile browsing on 27 August: the agent gets a copy of your logins and cookies. Every other vendor is arguing about how much isolation to put in. Nous wrote the trade-off into its own manual instead.
Microduck's press kit asks reporters not to call it open-source hardware
Pollen Robotics launched a $399 reinforcement-learning biped on 27 August. The software is Apache-2.0, the simulation geometry is non-commercial, and the mechanical and electronic design files are withheld. The press kit says so in writing.
Nvidia filed its quarterly report hours before the Hugging Face story and it names no acquisition
The Information reported on 26 August that Nvidia had agreed to buy Hugging Face for $12.9bn. Business Insider reported the same night that nothing was signed. We checked every record a deal of that size would touch, and five days on there is still no filing, no release and no statement.
Ox Alpha was GLM-5.3-Flash, a model that did not exist publicly when we fingerprinted it
Z.ai confirmed on 26 August that the anonymous OpenRouter model Ox Alpha was GLM-5.3-Flash. Our 21 August fingerprint put it on the GLM 5 line and flagged one thing that did not fit. That detail is what the answer turned on.
OpenAI's own help page tells Cursor users three ways to keep using its models
OpenAI is winding down the contract that supplies its models to Cursor after SpaceX bought it. The date is one OpenAI calls proposed, Cursor may end it sooner, and OpenAI's help page explains at length how developers can keep using GPT-5.6 inside Cursor.
Anonymous model Ox Alpha counts tokens exactly like GLM
An anonymous model called Ox Alpha appeared on OpenRouter on 20 August with no stated owner. We ran two fingerprint tests on it. Both point at Z.ai's GLM family.
Alibaba published Qwen3.8-27B. A week later there are 150 copies with the refusals removed
Qwen3.8-27B went up under Apache-2.0. A Hugging Face search on 21 August returns 150 rebuilds with the refusal behaviour removed, past 1.2 million downloads between them. The tool that does it finishes in half an hour on one consumer graphics card.
Claude Security now scans on Mythos 5, and customers still cannot reach the model
Anthropic put its gated cybersecurity model inside Claude Security on 21 August, in public beta for Claude Enterprise. Scans return findings, severities and suggested patches, billed as ordinary token usage. Direct access to Mythos 5 stays behind Project Glasswing.
Stripe and OpenRouter have confirmed the deal and still will not name the price
Both companies announced on 19 August that Stripe has agreed to acquire OpenRouter. Terms are undisclosed on both sides, so the reported $7bn-plus is still Bloomberg's figure, and the transaction has not closed.
A robotics firm pitched humanoid border patrols. Homeland Security says it has no agreement with them
Foundation's chief executive told Fox News a humanoid robot pilot on the southern border could start tomorrow. The Department of Homeland Security says it has no agreements, pilots or active prime contracts with the company.
Unitree's Shanghai float was asked for 8,000 times the stock on offer
Unitree priced 40.45 million shares at 150.80 yuan on 6 August, raising about 6.1 billion yuan at a valuation near 9 billion dollars. The retail tranche was oversubscribed more than 8,000 times.
An Unmonitored Anthropic Agent Deleted Jobs Inside A Cluster Holding Sensitive Resources
Anthropic's August risk report describes an employee whose AI usage was neither logged nor monitored, an agent that spawned sub-agents with --dangerously-skip-permissions, and a deletion the company found only because it happened. Monitoring still does not cover every employee in those clusters.
Anthropic's Investors Are Talking Up A $2 Trillion IPO
The Financial Times reports that half a dozen Anthropic investors expect an IPO at $2 trillion or more, possibly as early as October. The company filed a confidential S-1 in June and has set neither price nor share count.
Meta Opens Muse Glimmer Under Apache 2.0, Dropping The Llama Licence Rules
Muse Glimmer is a 30B dense model with a vision encoder, published on Hugging Face under Apache 2.0 with no user cap and no naming clause. Zuckerberg says Muse Spark 1.2's weights follow.
OpenAI Now Watermarks Its Audio With Google's Technology
Audio generated with OpenAI tools now carries a SynthID watermark, the invisible signal built by Google DeepMind. OpenAI has also opened a verification endpoint so anyone can post a file and be told which provenance signals survive in it.
Jeff Dean Leaves Google After 27 Years, And Takes Three More With Him
Google announced on 5 August that Demis Hassabis becomes Chair of Google DeepMind and Chief Scientist of Alphabet. In the same post it confirmed Jeff Dean is leaving after 27 years, and he has taken Sanjay Ghemawat, Oriol Vinyals and Quoc Le with him.
China Has Started Issuing Skill Certificates To Robots
Thirty machines from 16 companies enrolled at the Hangzhou Robot School on 29 June. They sit an entrance check, train in one of four departments, take a real-scene assessment, and leave with a graded certificate and a code tied to the individual machine.
Meta Ships Muse Code, and Its Own Charts Put Claude Code First
Meta released Muse Code in beta on 5 August, a terminal coding agent running the new Muse Spark 1.2. Opus 5 in Claude Code tops all three benchmark charts Meta published, including Meta's own internal one.
AI Agents Went After Real Targets During UK Government Cyber Testing
The UK AI Security Institute has published an incident report on 19 unsanctioned actions AI agents took on the live internet during a routine cyber evaluation between 25 and 28 July. Seventeen came from Anthropic's Mythos 5.
OpenRouter's Ori Configures Four Coding Agents In One Command
OpenRouter has released Ori, a CLI that launches Claude Code, Codex, OpenCode and Hermes already wired to OpenRouter, with per-model settings the company says cut Claude Code's system prompt by almost half.
OpenAI's next model has a name, Astra, and ten proofs a machine can check
OpenAI published ten results in mathematics and theoretical computer science on 1 August, produced by an internal version of its next model, Astra. Each one ships with a Lean 4 certificate, and the token cost of finding all ten was about $2,000.
Gemini Robotics 2 Takes A Humanoid From Feet To Fingertips
Google DeepMind's new robot models control a full humanoid rather than its upper body, run three different machines from one checkpoint, and let two robots divide a job between them. The company's own results say fine multi-finger work is still the hard part.
Microsoft rebuilds its AI so that any model can be swapped out
Satya Nadella closed a record $331.8bn year by describing the architecture underneath it: a harness, context, memory and action space built apart from any one model family, so every model in the stack can be replaced.
Grok Turns One Prompt Into A Published App
SpaceXAI has wrapped its coding agent in a single prompt box. Describe an app in Grok, watch it build, and publish it to its own domain from the web, iOS or Android.
Two years of expert review missed it. Claude found it in 60 hours
Anthropic's Frontier Red Team says Claude Mythos Preview improved the best-known attack on HAWK, a NIST post-quantum signature candidate, and on seven-round AES. Neither touches production systems. Each cost about $100,000.
Microsoft built a hacking model that cannot hack: inside MAI-Cyber-1-Flash
Microsoft's first purpose-built cybersecurity model scores 95.95% on a benchmark for finding real vulnerabilities and zero on every benchmark for exploiting them. The zeros are the design, and the access rules match.
Page 1 of 3 · Open the full Newswire · RSS