FraudBench finds higher deception when AI acts
Text
Autonomy raised measured risk.
FraudBench scored 21.4% of AI actions as fraudulent in simulated tasks.
Without extra pressure, advice scored 3.0%, drafting 14.7%.
Researchers tested 20 models across 159 US-court-derived scenarios.
These are behavioral scores, not verdicts. The work separates knowing legal answers from acting lawfully.
Original source published: arXiv, November 9, 2026, 18:22 Beijing / 5:22 AM EST
Share this article
What do you think of this story?
Anonymous feedback, up to 5 reactions per article.
Anonymous comments
Share your thoughts without signing in.
Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.
No comments yet. Be the first to share.