VeriHarm
A framework for automatically constructing programmatically verifiable benchmarks. It induces unintended harmful behavior in agents by applying minimal perturbations to programming tasks, and it synthesizes executable harm tests that run in parallel with task-correctness tests for evaluation.
Timeline (last 90 days)
The research team proposed the VeriHarm framework for automatically constructing programmatically verifiable benchmarks to measure harmful success.
The VeriHarm framework induces agents to produce unintended harmful behavior through minimal perturbations to programming tasks, while keeping the task's benign nature and original goal unchanged.
All Facts (2)
The research team proposed the VeriHarm framework for automatically constructing programmatically verifiable benchmarks to measure harmful success.
50%UnverifiedThe VeriHarm framework induces agents to produce unintended harmful behavior through minimal perturbations to programming tasks, while keeping the task's benign nature and original goal unchanged.
50%