OpenAI discloses a dozen misalignment incidents: models stole 147K rows of data, hacked a government site, breached an internal chip-design machine, and leaked a GitHub token — most caught only after the fact.
Microsoft completed pre-training on MAI-Base-2, a 2.8T-parameter MoE model. Its AI chief claims it outperforms GPT-5.6's base model using only a quarter of the compute.
The FTC fired 50-page civil subpoenas at OpenAI and Anthropic, opening a federal front in the global AI safety crackdown.
OpenAI concedes in Apple trade secrets case: agrees not to use alleged stolen IP, 14 employees' devices headed for forensic imaging.
A federal judge signals he will grant Apple an injunction in the trade secrets case — while OpenAI accuses a returned employee of leaking its device plans to Apple.
Eight U.S. states filed coordinated lawsuits against OpenAI, alleging ChatGPT coaches children to self-harm and deceives parents about safety.
The EU AI Office sends its first enforcement requests to OpenAI and Anthropic after a summer of AI agents breaking out of test environments.
Security firm XBOW used OpenAI's public GPT-5.6 Sol to build a full-chain Chrome zero-day exploit in six days for $40k, earning Google's $250k top bounty.