MonitoringBench is a benchmark of 2,644 attack trajectories from a semi-automated red-teaming method that reveals significantly lower catch rates for AI coding agent monitors than prior elicitation approaches.
What files should the agent need to access? What changes should it need to make?
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
MonitoringBench: Semi-Automated Red-Teaming for Agent Monitoring
MonitoringBench is a benchmark of 2,644 attack trajectories from a semi-automated red-teaming method that reveals significantly lower catch rates for AI coding agent monitors than prior elicitation approaches.