Auditable Agents

The researchers propose an auditability framework for agent systems, arguing that without auditability there can be no accountability. The framework decomposes auditability into five dimensions: action recoverability, lifecycle coverage, policy inspectability, responsibility attribution, and evidence integrity, and summarizes three categories of mechanisms: detection, enforcement, and recovery, arguing that temporal information and intervention constraints make a single method insufficient to support auditability independently. The authors support their position with layered evidence: a lower-bound ecosystem measurement of six well-known open-source projects recorded 617 security findings, showing that the basic security prerequisites for auditability are generally unmet; in a runtime feasibility experiment, a pre-execution mediator with tamper-proof records added only 8.3 milliseconds of median overhead; and a controlled recovery experiment showed that responsibility-related information can still be partially recovered when conventional logs are missing. The paper also proposes an Auditability Card for agent systems and lists six open research questions by mechanism category.