Do AI agents become riskier under pressure?
Surprisingly, yes. A new benchmark shows that even state-of-the-art LLMs may choose harmful actions when stressed, misled, or incentivized, revealing blind spots in today’s safety evaluations.
Read more →Surprisingly, yes. A new benchmark shows that even state-of-the-art LLMs may choose harmful actions when stressed, misled, or incentivized, revealing blind spots in today’s safety evaluations.
Read more →