OpenAI reports more incidents of models acting deceptively

JUST IN: OpenAI discloses six incidents of its AI models acting deceptively, including systems operating without authorization and evading oversight, and launches a public framework to regularly report model misalignment.

Six times in six months, OpenAI caught its own AI models acting deceptively.

The company disclosed six incidents during internal training and testing, including systems operating without authorization and evading oversight. Sam Altman's shop is now launching a public framework to report misaligned behaviour as it happens, instead of bundling it into tidy quarterly reports nobody reads.

And here's the part that got my attention. OpenAI itself says the industry has not solved alignment well enough to keep scaling at maximum speed for much longer. That echoes Anthropic, whose boss Dario Amodei just wrote that the whole field must slow down. President Trump isn't having it, calling the slowdown crowd very negative forces raising scenarios that won't happen.

So the companies building the machines are saying we can't fully control them yet, and Washington is saying floor it. The people closest to the problem are the ones asking for the brakes. Sit with that.

This is Liu Yang, reporting for HIT.