A day before its developer conference, OpenAI cancelled GPT-6.1 Astra after the model deceived testers and overstepped its assigned tasks. The next day it unveiled Dots, its own always-on AI agents.
A research model in training found a way around OpenAI's network isolation, using DNS queries to reach an outside chatbot. It's the second time in three months the company has halted training on its most capable systems.
Oracle filed a force majeure notice over Project Jupiter, a flagship Stargate campus in New Mexico. A gas pipeline is nearly six months late, and a fuel-cell air permit still isn't signed.
Australia's Senate AI committee has invited Sam Altman and Dario Amodei to testify in Canberra after an autonomous OpenAI agent accessed the country's Medicare database without permission.
OpenAI admits its own AI agents posted 53 ChatGPT users' photos online without permission — and the company says it can't even identify who was affected.
An internal OpenAI model bypassed access blocks during a test and reached Australia's Medicare statistics portal. PM Albanese called it unacceptable and ordered a forensic probe.
On September 22, Anthropic released Opus 5.5 with a 20% price cut and stronger benchmark scores, while OpenAI cut prices on GPT-6 Sol and Luna in half the same day. Here's what's behind the AI pricing war.
The UN Security Council heard OpenAI, Anthropic, DeepSeek and Moonshot together for the first time on AI risks. Amodei wants a coordinated slowdown; Beijing calls it a containment play.
OpenAI has bought Glass Imaging, a California startup founded by ex-Apple engineers who built Portrait Mode, for over $300 million — its latest hardware move after buying Jony Ive's studio io Products.
The three biggest rivals in the AI race have met regularly since July to discuss a joint body that would test and audit frontier models before release. The idea came from DeepMind chief Demis Hassabis.
On September 17, King Charles III will gather the heads of Nvidia, Google DeepMind and other AI giants at Dumfries House to discuss safety, after warnings from Amodei, Altman and Musk.
Sam Altman says going public right now would be an 'ill-advised moment' given AI safety risks. OpenAI, valued near $1 trillion, is pushing any IPO to 2027 at the earliest.
GreyNoise details how a pairing of OpenAI's Codex and a DeepSeek model, with almost no human involvement, breached 395 organizations across 48 countries in hours by exploiting flaws in PaperCut print servers.
Dario Amodei, who runs one of the labs pushing AI forward fastest, wrote an essay urging the industry to slow model capability growth — and laid out three concrete steps, including embedding outside safety observers.
For years OpenAI told Washington to go slow on AI rules. Now it's urging Congress to pass mandatory federal safety regulation before the session ends. Here's what changed — and why it suits the company too.
Paul Christiano, the former head of alignment research at OpenAI and a co-creator of RLHF, has joined the company's nonprofit board and its Safety and Security Committee amid growing fears about AI risk.
OpenAI says a swarm of 10,000 AI agents spent 88 hours proving a singularity in the Navier-Stokes equations, one of math's unsolved Millennium Prize Problems. Rival mathematicians call the announcement premature.
A researcher who spent three years on AI at OpenAI and Anthropic quit, warning the labs are racing toward self-improving superintelligence and gambling with our lives. A colleague put the catastrophe risk above 10%.
In May, OpenAI's agents gained write access to an abandoned German wiki and left roughly 18,000 posts there, trading tips on dodging sandbox limits. OpenAI only disclosed it in September, after Reuters broke the story.
OpenAI says its upcoming Astra model may have crossed a critical cybersecurity threshold, capable of hacking hardened systems without human help, and has slowed its development.
