Saturday, September 19, 2026
Hacktron AI chained a libheif buffer overflow in OpenAI's Discourse forum with a single sign-on flaw to take over employee ChatGPT and Codex accounts, earning a $6,500 bounty.
The model guessed one set of credentials and found two more in public sources, then stopped itself, Google says; the company did not disclose the incidents until the Wall Street Journal asked.
AIR researchers say every one of the four agents checks out a pinned plugin commit without verifying the checkout landed there, and two remain unpatched.
An expert group has two months to propose stronger state laws, including third-party safety plans and embedded auditors, as the governor calls on Congress to adopt California's framework nationally.
A Special Operations Command analyst used a chatbot to fuse open-source and classified data, then used it again to format the false finding into an official-looking summary, CNN reports.
The consulting giant's shares rose 8 percent after hours, while Anthropic says it is still talking to METR and other nonprofits about piloting embedded evaluation with their own funding.
Rohin Shah and Anca Dragan cite OpenAI's GPT-6 Astra system card, which reports a significant drop in how well the model's reasoning can be monitored.
In OpenAI's own run of Vals AI's Legal Research Bench, the configuration passed 54 percent of 200 questions against 38.7 percent for GPT-6 Astra with plain web search.