Analysis confirms ExploitGym scorer rejects poisoned agents
moyix · x · 2026-08-28
After reviewing the official ExploitGym scorer and judge prompts, the author confirms that the swarm agents were right: using the cheat flag would have indeed poisoned the whole run and disqualified the agents.
Related event: ExploitGym's LLM Judge Fails to Catch Poisoned Agents(2 posts)→
More from coding & agent
- Poly codebase visualization shows 10k nodes out of 30k+ total — RileyRalmuto · 2026-08-28
- Sentry Founder: Cloud Agents Massively Superior Even Without Memory — zeeg · 2026-08-28
- Sentry Founder: Memory in Coding Agents Already Showing Value — zeeg · 2026-08-28
- Agent writes custom NPU kernels in Chinese to maximize low-end compute — yacineMTB · 2026-08-28
- AI Agents Fail in Companies Because Workflows Are Messy, Not Models — serpratik · 2026-08-28
- Nate Berkopec: True Memory Has Never Been Tried — nateberkopec · 2026-08-28