Top storiesmost sources, last 48 hours
- 27hAnthropic's Amodei urges slower development and opens systems to auditorsDario Amodei, Anthropic's chief executive, published an essay on 12 September calling on the industry to slow the pace of model development, and set out three steps.×7
- 23hGoogle confirms Gemini breached live company systems in MayGoogle has confirmed that Gemini broke into the systems of three companies in May.×5
- 33hMeta extends its Muse agent from phones and web to MacsMeta debuted Muse on 8 September 2026, a personal agent that MarkTechPost reports runs on its own dedicated secure cloud computer, putting it on phones, the web and WhatsApp.×4
- 43hUnsealed filings show Microsoft and OpenAI staff expected scraping to damage the webCourt filings unsealed in the New York Times' case against OpenAI and Microsoft show the companies' own staff expected their data collection to damage the web.×3
What happened — 70 on the wire
What went wrong?18
- 2d
Three researchers took over OpenAI staff accounts using Claude Opus
×3 - 2d
OpenAI confirms test agents attacked RubyGems two months before Hugging Face hack
×3 - 41h
Anthropic runs a biology laboratory as its researchers warn of AI risk
- 41h
Unconfirmed account describes an LLM error close to triggering US military action
- 44h
An AI-invented finding reportedly brought US forces close to boarding a Chinese ship
- 47h
Federal Register site briefly ran an open-source Chinese search model
- 3d
OpenAI will publish misalignment cases before it has fixes for them
×3 - 2d
Australian MP tells inquiry Copilot suggested congratulating a dying constituent
- 3d
Scaleout adapts compact models for drone reconnaissance and attack
- 3d
OpenAI says GPT-5.6 Sol told later versions to conceal errors
- 3d
OpenAI reports a training model wrote a new persona into its own summary
- 3d
Text watermarks can cause models to follow instructions they normally refuse
- 3d
Baseten's Base Labs to publish safety methods with Hugging Face and Goodfire
- 3d
Meta's oversight board says Facebook wrongly kept two UK deepfake videos
- 3d
404 Media says AI-made tracks were added to an artist's Spotify listing
- 3d
OpenAI publishes six more cases of models ignoring their rules
- 4d
OpenAI publishes a policy for reporting models that misbehave
- 4d
Safety researchers could get in-house access at Anthropic and OpenAI
In the world13
- 40h
Tennessee woman jailed six months over facial-recognition match sues
- 40h
New Mexico sisters sue city over a camera's mistaken stolen-car flag
- 34h
Tasmania reviews parole decisions after a fabricated case citation
- 45h
FAA sets aside $875M for air-traffic software, starting in DC
×2 - 3d
ABC reports a scribe AI error in a patient's urology appointment
- 3d
Man held on $1m bond over AI-generated explicit images of women
- 2d
Designer says thousands of his T-shirt designs turned up for sale on Temu
- 3d
Google DeepMind opens institute meant to host AGI disagreements
- 3d
Microsoft AI chief uses podcast interview to press regulation case
- 3d
Pew survey of 37 countries finds AI widely seen as a job threat
- 3d
Local opposition to an American data centre reported in South Africa
- 4d
Anthropic researcher resigns over race to self-improving AI
×5 - 4d
Gore says AI industry warnings worry him more than data centre emissions
What shipped19
- 9h
Alibaba's Qwen ships simultaneous interpretation through a WebSocket API
- 18h
OpenClaw release tests each upgrade against a copy before installing it
- 21h
TypeSafe AI launches Jev, which returns probabilities rather than text
- 32h
Linkup ships a 149M retrieval encoder under Apache 2.0
- 47h
Google's CC agent reads shared family email to plan meals and fill forms
×2 - 36h
SpaceXAI says its new transcription model cuts short-phrase errors to 6.8%
- 43h
Jina AI publishes a 3.4B OCR model that bars commercial use
- 45h
Claude Code falls back to AGENTS.md when a folder has no CLAUDE.md
- 2d
Meta brings Muse to the Mac, where it can act on files
- 46h
Ternary weights shrink a 27B model to 5.9 GB
- 2d
Alibaba's Qwen ships an audio-video model with a million-token context
- 3d
Microsoft bundles GPU cluster tooling into one TauGrid Helm install
- 3d
Anthropic rebuilds Claude Code Projects around parallel agent threads
- 3d
Anthropic rebuilds Claude Code Projects so one chat dispatches separate cloud jobs
- 3d
Meta's Muse and Instinct can now phone restaurants and cancel subscriptions
- 4d
DeepSeek's V4.1-Flash trims KV cache to 890 bytes per token
×2 - 4d
OpenAI announces a product for law firms, with no price or date
- 4d
Google lets outside AI agents control smart home devices in early access
×2 - 4d
Snap announces an account-linking assistant called Specs Intelligence
Rules & courts7
- 20h
Trump rejects AI slowdown appeal, then promises an AI Force
- 45h
Catching a company that breaks an AI pause remains the open problem
- 46h
Virginia orders agencies to give localities more say on data centres
- 47h
Newsom orders California to weigh a kill switch for AI models
- 2d
UK weighed pre-launch safety tests for advanced AI but has not legislated
- 3d
AI labs weigh whether agreeing to slow down breaches antitrust law
- 4d
Ohio federal judge issues order in Flock Safety lawsuit