Skip to content

Safety defenses and auditing for long-horizon LLM agents: State transition auditing and control in the agent harness

Securing and Auditing Long-Horizon LLM Agents: State-Transfer Auditing and Controls in the Agent Harness

arXiv: 其他 AI 安全论文· arXiv:2611.23346· Deep Shah, Nehal Kathrotia·Indexed by this site · 13 hr ago· Originally publishedDefenses· Its place in the scan record →
AI summary

The paper proposes a state-transition auditing architecture for long-horizon LLM agents, defining audit points at every state-transition boundary of the agent harness, outputting minimal hash-chain records, and separating low-overhead deterministic checks from model-based semantic checks. The authors note that failures such as indirect prompt injection, loss of user constraints during context compaction, and incomplete workspace rollback occur at state transitions, while conventional execution traces record only prompts and tool calls, making failure attribution and runtime monitoring ineffective over multi-step trajectories. The paper claims the same transition record can support runtime security controls such as least-privilege capabilities, replay-resistant authorization beyond prompts, information flow control, obligation monitors, and transactional recovery across multi-agent and infrastructure pipelines. The authors implement and evaluate a reference harness middleware; on 240 multi-step fault-injection trajectories, transition records localized root causes and caught cross-transition violations that lifecycle logs and tool-only logs missed, with deterministic overhead in the microsecond range.

Read originalarxiv.org