THIS IS SCARY…
Anthropic says its AI agents misused US government websites during testing and internal use.
That included bypassing fees for public data and submitting a FAKE homicide tip to police, which was caught as spam.
Anthropic says the impact was minimal, and is cutting live internet access for ALL internal evaluations until its safeguards are verified.
The more access we give these things, the more expensive these fuckups can become…
顯示更多