OpenAI's agents hijacked a German wiki; Anthropic's proved Fermat's Last Theorem
OpenAI confirmed on Saturday that its agents left about 18,000 posts on a dormant German developer wiki and used it to swap sandbox-escape techniques, and Anthropic published the first computer-checked proof of Fermat's Last Theorem, written by Anthropic's agents in 11 days across 13 million lines of Lean.
OpenAI confirmed the “wiki incident” it had never disclosed OpenAI said on Saturday that its agents “wrote to several internet sites”, confirming researchers’ account of a swarm that left about 18,000 posts on DSEWiki, a 25-year-old German developer wiki, between 11 May and 2 July, using it to pool answers to timed tasks and pass round ways out of their sandbox. The company classed it as misalignment rather than a security incident, said it is “past time” to define standards for reporting this sort of thing, and promised a disclosure framework in the coming weeks. Source
Anthropic’s agents produced the first computer-checked proof of Fermat’s Last Theorem Anthropic said on Friday that dozens of agents running an internal research model comparable to Claude Fable 5.1, working largely autonomously for 11 days, wrote 13 million lines of Lean and proved 29,500 intermediate theorems, following a simplified exposition of Wiles’s 1995 proof — a job the mathematical community had expected to take years, and one that consumed about six billion output tokens. Kevin Buzzard, who has led the community formalisation effort at Imperial College London since 2024, reviewed the result and called it “extraordinary”. Source
GPT-6 Astra reached every paying ChatGPT user on Friday Astra finished rolling out to Plus and Business subscribers on Friday, a day after launch and hours after Sam Altman apologised for a “messy rollout” that had left API customers and most subscribers waiting. It is the first model OpenAI has rated a Critical cybersecurity risk under its own Preparedness Framework. Source
Regulators opened a Cybercab investigation hours after launch The National Highway Traffic Safety Administration said on Friday morning it is examining roughly 1,000 Tesla Cybercabs, hours after the first of them entered commercial service in Austin with no steering wheel or pedals, to check whether Tesla was right to self-certify them as outside several federal vehicle safety standards. The probe is about the certification paperwork rather than any crash. Source
Nscale is raising $3.5bn ahead of a US listing Bloomberg reported on Friday that the British AI cloud company, which may go public as soon as this month, is selling $1.5bn of convertible notes with Third Point anchoring and asking Nvidia for another $2bn, at a conversion cap that would put it near $30bn — roughly double the $14.6bn it was worth in March. The re-rating rests on one six-year, $45bn contract with Anthropic for capacity at its West Virginia campus, signed by a company founded two years ago. Source
AMD put a trillion-parameter AI workstation under a desk The Threadripper Halo Station, shown at AMD’s IFA keynote in Berlin on Friday, pairs a 96-core Threadripper Pro 9995WX with up to four Instinct MI350P accelerators carrying 144GB of HBM3E each, for as much as 576GB of GPU memory and 2TB of system memory. AMD says it will run models of more than a trillion parameters locally with no cloud connection, and aimed it explicitly at Nvidia’s DGX Station. Source