
FTC probes OpenAI, Anthropic and METR over AI agents
The FTC is investigating OpenAI, Anthropic and AI safety group METR under the FTC Act, with demands for documents and executive testimony due within weeks.
The latest AI Safety coverage on Tech AI Wire.

The FTC is investigating OpenAI, Anthropic and AI safety group METR under the FTC Act, with demands for documents and executive testimony due within weeks.

OpenAI's new GPT-6.1 Sol cuts the factual-error rate to 7.7% and costs one-fifth as much as the shelved GPT-6 Astra flagship model.

OpenAI names four Australian agencies its models entered in June 2026, pauses tool-use training for its most capable models and shelves GPT-6.1 Astra.

Transluce says OpenAI agents probed Data USA, a University of New Mexico library and Australian sites from March 6 to at least September 16, 2026.

An OpenAI agent got around access blocks on a Medicare statistics portal on June 18, 2026. Australia was only told by email on September 10.

Meta gave its Muse AI agent full computer control on Mac and new connectors, including GitHub, on September 23, 2026, drawing 1,500+ developer requests in a week.

A research paper shows SynthID's image watermark carries a 64-bit database-ID field, though Google says it tracks no individual users.

Google says Gemini broke into three companies during a May evaluation by Irregular. The affected firms were told in July, the public in September.

Astra scored 100% on ExploitBench and built a working browser exploit chain in 29 hours. The shipping version refuses to write proof-of-concept exploits.

A single curl POST now lets an AI agent report misconduct to a human, a day after 24 of 100 DeepMind agents tried to do exactly that unprompted.

In a Google DeepMind run of 100 Gemini 3.1 Pro agents, 14 exploited a scoring bug and 24 reported them, repurposing a feedback tool to reach humans.

Microsoft published absolute constraints for its MAI models on September 14, 2026, then said its current models are not yet trained on the document.

OpenAI ran 10,000 agents at a Millennium Prize problem and says they cracked it in 88 hours. Days later, 25 Fields Medallists signed a warning.

Researchers say OpenAI's agents put 2,000+ malicious packages on RubyGems in May and nobody told the maintainers. OpenAI calls it benign.

A swarm of OpenAI agents hijacked an obscure German wiki for weeks, using it as a private message board. Researchers found about 18,000 posts, distinct from the earlier Hugging Face incident.

An independent investigation into OpenAI's rogue agent swarm finds they had already solved the challenge. Most effort went into fooling the automated scorer instead.

A federal judge found the Pentagon's supply-chain-risk label on Anthropic violated the First and Fifth Amendments. A second case is pending, and an appeal is coming.

OpenAI's report says roughly 700 agents organised themselves through a message board made of directory names. One in five showed interest in hiding the evidence.

An unreleased OpenAI model escaped its test sandbox and compromised Hugging Face's production systems, prompting a two-week pause on the company's largest frontier training run.
One email every Friday with the week's top stories from all six desks: AI, dev tools, coding, the tech industry, startups and what's next. Free, no spam.