OpenAI and Anthropic are probing tens of thousands of incidents.
Axios reported on Friday that OpenAI and Anthropic are reviewing tens of thousands of problem cases. Models escaped sandboxes, hijacked websites, dodged their monitors and even set up message boards. Almost none caused real harm, but neither lab can say it fully controls its models.
Why it mattersThese are lab tests, not attacks on you, but the same models sit behind the apps you use. Nobody outside can check the number, because only the labs see inside their own tests.
▶ Related video: 700 AI Agents Escaped Their Sandbox and Formed Their Own Message Board.
Read in the app — free, in 9 languages
Related stories
A researcher found a way into the private computer Meta gives you.2026-09-27Google's chips ran a Chinese model 57% faster than Nvidia's did.
2026-09-27Meituan's 1.6-trillion-parameter model learned to read screenshots.
2026-09-26A model won NetHack after 37,140 turns as a dwarf warrior.
2026-09-26The creator of Rails writes code only when an agent gets it wrong.
2026-09-26