Agent Monitoring

Agent Monitoring Services

Agent Monitoring Notion

Proposes an MRT framework that covers the situational-awareness levels of both the agent and the monitor, evasion strategies such as prompt injection, and the environments for tool use and computer use. In particular, it introduces a hybrid monitor scaffolding approach that combines hierarchical summarization with sequential inspection.

A "weak-to-strong" scaling effect where a weak model can effectively monitor a strong agent. Experiments find that "agent awareness" (whether the agent knows it is being monitored) is the dominant factor that substantially reduces the monitor’s reliability.

Reliable Weak-to-Strong Monitoring for LLMs

Explore how weak LLM monitors can reliably oversee stronger LLM agents, with findings on scalable oversight and safety in agentic AI systems.

https://labs.scale.com/papers/mrt

Reliable Weak-to-Strong Monitoring for LLMs

Reliable Weak-to-Strong Monitoring of LLM Agents

We stress test monitoring systems for detecting covert misbehavior in autonomous LLM agents (e.g., secretly sharing private information). To this end, we systematize a monitor red teaming (MRT)...

https://arxiv.org/abs/2508.19461

Monitor Red Teaming(MRT) benchmark

Systematically stress-test the reliability of monitoring systems at detecting covert misbehavior, such as leaking sensitive information or sabotaging the system.

Reliable Weak-to-Strong Monitoring of LLM Agents

We stress test monitoring systems for detecting covert misbehavior in autonomous LLM agents (e.g., secretly sharing private information). To this end, we systematize a monitor red teaming (MRT)...

https://arxiv.org/abs/2508.19461

SHADE Arena

arxiv.org

https://arxiv.org/pdf/2506.15740

별도 안전 분류 모델 없이 기존 KV cache를 재활용해 저비용으로 대규모 모니터링을 수행. Claude가 자신의 응답 직후 대화 전체의 위험성을 스스로 평가하도록 학습.

AI Introspection

Scaling Misuse Detection with Self-Assessment

Self-assessment lets Claude assess the safety of its own responses. It scales well with model capabilities and helps us identify emergent harms.