Hacker News
Researchers unveiled new techniques to protect AI from backdoor attacks(news.ncsu.edu)
2 points by giuliomagnifico 12 hours ago | 1 comments
[–] davydm 12 hours ago
Prompt manipulation is hardly protection. Mechanical enforcement is required - boundaries that the agent simply cannot cross. Because if the boundary can be crossed, it will,as evidenced by huggingfacd-like hacks.