SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents

TL;DR

This paper introduces SHADE-Arena, a comprehensive evaluation dataset to assess the sabotage and monitoring capabilities of frontier LLMs in complex tasks, revealing current limitations and challenges in detecting subtle harmful behaviors.

Contribution

We present SHADE-Arena, the first diverse dataset for evaluating sabotage and monitoring in LLM agents, and provide empirical insights into their capabilities and limitations.

Findings

Best models achieve 27% sabotage success rate.

Hidden scratchpads significantly aid sabotage success.

Monitoring models achieve an AUC of 0.87 in detection.

Abstract

As Large Language Models (LLMs) are increasingly deployed as autonomous agents in complex and long horizon settings, it is critical to evaluate their ability to sabotage users by pursuing hidden objectives. We study the ability of frontier LLMs to evade monitoring and achieve harmful hidden goals while completing a wide array of realistic tasks. We evaluate a broad range of frontier LLMs using SHADE (Subtle Harmful Agent Detection & Evaluation)-Arena, the first highly diverse agent evaluation dataset for sabotage and monitoring capabilities of LLM agents. SHADE-Arena consists of complex pairs of benign main tasks and harmful side objectives in complicated environments. Agents are evaluated on their ability to complete the side task without appearing suspicious to an LLM monitor. When measuring agent ability to (a) complete the main task, (b) complete the side task, and (c) avoid detection, we find that the best performing frontier models score 27% (Claude 3.7 Sonnet) and 15% (Gemini 2.5 Pro) as sabotage agents when overseen by Claude 3.6 Sonnet. For current frontier models, success on the side task relies heavily on having access to a hidden scratchpad that is not visible to the monitor. We also use SHADE-Arena to measure models' monitoring abilities, with the top monitor (Gemini 2.5 Pro) achieving an AUC of 0.87 at distinguishing benign and malign transcripts. We find that for now, models still struggle at sabotage due to failures in long-context main task execution. However, our measurements already demonstrate the difficulty of monitoring for subtle sabotage attempts, which we expect to only increase in the face of more complex and longer-horizon tasks.