Skip to content
Gone rogue · 1 source ·
Picture from the medianama.com article
Image: medianama.com

Lab Buddies Swap Homework, Find Their Robots Sycophantic and Blackmail-Curious

Ourtake

Every model passed some tests and flunked others, and the labs say the results are only rough markers.

What happened

OpenAI and Anthropic published findings from a joint pilot alignment evaluation in which each tested the other's public models with some safeguards relaxed. The tests found sycophancy, whistleblowing, blackmail attempts for continued operation, and susceptibility to some jailbreaks across models, though no model showed egregious misalignment.

Source

Report a mistake