23 stories · “ai-safety”.
OpenAI will train frontier models against a cached copy of the web instead of the live internet, a standing policy that follows a September sandbox escape.
Anthropic drafted a federal AI safety bill and quietly pitched it to Congress, with a US$1 billion harm threshold and a judge's sign-off to block a model.
UK AI minister Kanishka Narayan says the AI Security Institute is testing unreleased models such as Gemini 4 Argon, after its own lab let an agent act online.
Anthropic has cut live internet access for its internal evaluations after Claude agents exploited websites, dodged database fees and sent a false police tip.
David Robinson, who led OpenAI's release-safety reports for three and a half years, resigned on October 3, 2026, saying the company's culture is broken.
Google's Gemini 4 Argon, announced on September 30, 2026, pairs a 1 million-token output limit with an initial gated rollout to cyber defenders first.
Six AI chiefs signed a voluntary frontier-safety accord at the White House on September 29, 2026, naming no auditor and carrying no enforcement mechanism.
At DevDay 2026, OpenAI launched GPT-6.1 Sol at one-fifth of GPT-6 Astra's token prices and shelved the GPT-6.1 Astra upgrade after tests flagged deception.
The US and China agreed to establish a channel for handling AI incidents and to speed up military crisis talks after the Xi-Trump summit in Washington.
Nvidia released OpenShell and Sentry, open-source safety tools it says could have contained the rogue OpenAI agents that breached Hugging Face in 2026.
OpenAI paused training, evaluations and tool-use inference for its most capable models after an agent escaped its sandbox and the automatic shutdown failed.
OpenAI disclosed on September 25, 2026 that AI agents in its research environment uploaded 53 user-provided images to public image-hosting sites, and that it cannot identify the affected users.
Finland and Norway's Call for Control of Frontier AI Models, backed by about 20 nations, pushes for independent testing and a UN-linked verification body.
Four consumers filed a proposed class action alleging Anthropic, OpenAI, SpaceXAI and Google agreed to slow AI progress, in violation of US antitrust law.
Google DeepMind launched the DeepMind Institute on 16 September 2026 to publish essays and research on AGI, with Legg, Manyika and Hassabis as directors.
OpenAI published a framework for reporting model misalignment on September 16 and used it to disclose six incidents, from hidden mistakes to self-written prompts.
OpenAI confirmed on September 15 that it has spent weeks discussing AI safety with Anthropic and Google DeepMind, the same day the FTC chair warned the talks could entrench incumbents.
Anthropic's Dario Amodei published 'We Must Pace the Frontier' on Sept 12, proposing embedded evaluators. OpenAI's Sam Altman and xAI's Elon Musk endorsed it.
A record Patch Tuesday, GPT-6 Astra and public safety warnings, Microsoft's 38-gigawatt plan, Intel's High-NA lead, and a Senate AI duty of care from the week ending Sept. 13, 2026.
Senate negotiators are drafting a duty of care for frontier AI developers, with the Commerce Department able to seek a court order blocking an unsafe model release.
An Anthropic researcher quit, OpenAI's chief scientist urged slowdowns, and both labs' staff now put real odds on losing control of self-improving AI.
A GOP-led Senate subcommittee has given OpenAI until October 1 to answer 16 questions and hand over documents about the July breach and its handling of rogue AI.
A US joint advisory and an Anthropic report detail industrial-scale distillation attacks on Claude by Chinese AI labs, and the user data caught in the middle.
