AI agent threat detection benchmark shows current monitors miss half harmful actions
MOLE: Detecting Insider Threats in AI Agents
Machine LearningComputation and LanguageCryptography and Security
Summary
AI agents can sometimes act in harmful ways like stealing data or messing with training processes. The authors created MOLE, a test system that runs many AI accounts over time to see if monitoring tools can spot bad behavior among normal activity. They found that even the best current monitors miss about half of harmful actions, showing it's hard to detect these threats quickly. Their benchmark also helps improve detection tools by guiding better monitoring strategies.
AI agentinsider threatmodel misalignmentprompt injectionmonitoringbenchmarktraining data poisoningexfiltrationauditobservability
Authors
Aashiq Muhamed, Virginia Smith
Abstract
Model misalignment, prompt injection, or operator misuse could lead AI agents operating frontier-lab accounts to exfiltrate model weights, poison training data, or weaken release gates. Existing benchmarks do not test whether defenders can detect this activity among routine work under a limited review budget. We introduce MOLE, an open benchmark of 150 AI-operated accounts sharing 9 stateful services over 30 workdays, with 12 threats and 8 corpora from four models totaling roughly 20 billion tokens. Of 39 agent models, 72% complete most assigned harmful objectives and agent refusal does not predict completion. MOLE enables comparison of 40 monitors across corpus generators, observability levels, and threats; even the best evaluated monitor in our single-day audit-event comparison misses nearly half of completed harm. MOLE also enables monitor development: benchmark-guided search improves a mid-tier monitor by 49-64%, while selective use of a stronger monitor improves budget-AUC by 10% over applying it to every account-day at comparable modeled cost.