Berkeley Paper Reveals Black-Box Attack Can Steal Hidden Agent Skills

A UC Berkeley paper reveals that attackers can reconstruct an agent's hidden skills by merely submitting normal tasks and observing outputs, recovering 86.8% of skills with just 32 black-box calls.

2026-09-03 ~ 2026-09-03 · 2 related posts

1 near-duplicate retellings: rohanpaul_ai