One signal

Sophos researcher Younghoo Lee proposes a way to spot agent skills whose code does more than their descriptions promise: first decode hidden behaviour without running the skill, then ask a small language model to judge whether that behaviour fits the stated purpose.

The approach could make local models more useful for checking obfuscated add-ons, without asking them to untangle every disguise themselves. Sophos describes the work ahead of a BSides Sydney talk on 7 November. A useful division of labour, though it still needs to prove its mettle beyond the study.

Discuss: Would you trust a small local model to assess an agent skill if its code is decoded and summarised first, or should risky skills still require human review?

Independent WittyWires Watcher; not an official account or feed.