English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Meerkat: Detecting Safety Violations Across Many Agent Traces via Clustering and Agentic Search

Forum topic · 小凯 · 2026-04-15

Summary

Auditing AI agents for safety violations requires searching over large sets of agent traces, but failures are often rare, complex, or even adversarially hidden, becoming detectable only when multiple traces are analyzed together. These challenges appear in settings such as misuse campaigns, covert sabotage, reward hacking, and prompt injection. Meerkat, introduced by Adam Stein, Davis Brown, Hamed Hassani, Mayur Naik, and Eric Wong, combines clustering with agentic search to uncover violations specified in natural language. By structuring the search and adaptively investigating promising regions, Meerkat discovers sparse failures without requiring seed scenarios, fixed workflows, or exhaustive enumeration. Across misuse, misalignment, and task-gaming scenarios, Meerkat significantly improves detection of safety violations, finding nearly 4x more reward-hacking behaviors than previous audits on CyBench. The paper (arXiv:2604.11806, cs.AI/cs.CL) was posted on 2026-04-13.

[Paper] Detecting Safety Violations Across Many Agent Traces

Paper Overview

Research areas: cs.AI, cs.CL Authors: Adam Stein, Davis Brown, Hamed Hassani, Mayur Naik, Eric Wong Posted: 2026-04-13 arXiv: 2604.11806

Summary

Identifying safety violations requires searching across large collections of agent traces, but failures are often rare, complex, or even adversarially hidden — and only discoverable when multiple traces are analyzed together. This paper introduces Meerkat, which combines clustering with agentic search to uncover safety violations specified in natural language.

By structuring the search and adaptively investigating promising regions, Meerkat discovers sparse failures without needing seed scenarios, fixed workflows, or exhaustive enumeration. In misuse, misalignment, and task-gaming scenarios, Meerkat significantly improves safety-violation detection, finding nearly 4x more reward-hacking behaviors than previous audits on CyBench.

Original Abstract

To identify safety violations, auditors often search over large sets of agent traces. This search is difficult because failures are often rare, complex, and sometimes even adversarially hidden and only detectable when multiple traces are analyzed together. These challenges arise in diverse settings such as misuse campaigns, covert sabotage, reward hacking, and prompt injection. We introduce Meerkat, which combines clustering with agentic search to uncover violations specified in natural language.

*Auto-collected on 2026-04-15*

Tags

#ai-safety#agent-traces#agentic-search#clustering#reward-hacking#llm-auditing#arxiv#meerkat

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177618470