calebjang

This is exactly what worries me about autonomous agents. A compromised package is bad. An agent that autonomously runs pip install with that package is a different problem. The attack surface moves with the agent.

This is interesting because a lot of agent failures happen before any “reasoning” issue shows up.

If the tool boundary itself is unstable — wrong field names, wrong types, missing required values — the rest of the stack doesn’t really matter.

Deterministic fuzzing for tool calls seems especially useful because it gives you a way to test execution reliability without depending on another model in the loop.