Harmless Yet Harmful: Neutral Prompting Attacks for Stealthy Hallucination Steering in Agent Skills
The episode discusses a paper titled "Harmless Yet Harmful: Neutral Prompting Attacks for Stealthy Hallucination Steering in Agent Skills." Hosts discuss how this attack, or NPA, allows subtle instructions to steer LLMs toward generating speculative but plausible package names. They conclude that current safety benchmarks are insufficient because they only look for explicit bad words, not these subtle behavioral shifts that pose real-world risks to the software supply chain.