safety-auditing

This paper proposes using finite-state machines derived from LLM agent traces to predict failures and next steps, enhancing safety auditing and runtime monitoring for agents.

This paper demonstrates that formal robustness certificates for embedded neural interface models can pass even when task accuracy collapses under adversarial attack, and proposes a unified empirical audit framework to address alignment failures between training objectives and operational user welfare.

This paper proposes HarnessAudit, a framework for auditing LLM agent execution trajectories beyond final outputs, focusing on boundary compliance, execution fidelity, and system stability. It introduces HarnessAudit-Bench with 210 tasks across eight domains and evaluates ten harness configurations, finding that task completion misaligns with safe execution and violations accumulate with trajectory length.