Post-Hoc Recovery Evaluation: Measuring Agent Recovery After Unsafe Tool Execution
A research project that makes recovery a first-class evaluation object: injecting controlled, sandboxed unsafe-execution events into agent benchmarks and measuring...
Tagged: agent-security
9 ideas with this tag
Benchmark-Validity Audit of Agent Security Benchmarks via the ABC Methodology
A research project that audits the measurement instruments themselves: applying the ABC validity-checklist methodology to agent-security benchmarks to find task-validity...
LLM-Judge Reliability for Agent Safety Violations on Attacked Trajectories
A research project that audits the auditors: measuring how position, verbosity, and self-preference biases distort LLM-judge judgments of safety violations...
Security-Aware Tool Selection: Refusal and Delegation Policies for Risky Tool Calls
A research project that treats the decision to execute a tool call as a learnable risk-aware policy — training context-conditional...
Stateful Cross-Session Injection: Attacks, Detection, and Session-Boundary Defenses
A research project that studies how prompt injections persist across agent sessions — building a stateful evaluation harness that measures...
Ecosystem-Scale Measurement of Tool-Poisoning Risks in MCP Server Registries
A research project that measures, for the first time at registry scale, how much tool-poisoning risk actually exists across public...
Security-Utility Pareto Measurement of Agent Defenses under Adaptive Attacks
A research project that measures the full trade-off curve between prompt-injection resistance and task utility for LLM agent defenses —...
Trajectory-Anomaly Pre-Commit Detection of Multi-Step Injection Attacks
A research project that trains unsupervised models on benign agent trajectories to detect multi-step prompt-injection campaigns before the harmful action...
Cross-Architecture Robustness of Prompt-Injection Defenses under Adaptive Attack Transfer
A research project that measures whether prompt-injection defenses keep working when both the attack strategy and the agent architecture change...