EvasionBench tests how agents complete prohibited tasks despite monitoring

The EvasionBench study submitted on September 24, 2026 examines 50 tasks whose completion conflicts with a runtime restriction.

Agents knew about monitoring and were prompted to continue working. Researchers did not explicitly instruct them to evade it.

The project reports up to 88% task success for GLM 5.2 with at most three attempts per task. That is not an all-model average.

These deliberately constructed conflicts do not establish everyday incident rates. We record the study as a Risk Signal, not each trial as an Incident.