StealthyPoison corrupts AI skills without breaking their execution
Text
Procedural poisoning can survive package screening.
StealthyPoison poisoned AI skills while leaving their tasks running normally.
Researchers evaluated altered procedures across 566 skills.
GPT-5.4-mini detected 45.8% of the poisoned packages in screening tests.
The controlled study separates spotting malicious instructions from checking whether a procedure is correct.
Original source published: arXiv, November 9, 2026, 18:05 Beijing / 5:05 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.