Identity Control Arena
ControlArena setting for identity-delegation attacks, built at the AI Control hackathon in Mar 2026.
Highlights
- Team lead; report: Does structured identity context improve LLM-based monitoring? It depends on calibration
- 48-hour hackathon with Apart Research and Redwood Research
Overview
Identity Control Arena is an evaluation environment built on the ControlArena framework to study identity delegation attacks in AI systems. The project was developed during the 48-hour AI Control Hackathon 2026, organized by Apart Research and Redwood Research.
Key Results
- Format mismatch discovery: A critical finding where format inconsistencies between attacker and monitor silently zeroed out all monitoring scores, highlighting a subtle failure mode in AI control evaluations
- Behavioral monitoring: The implemented behavioral monitor successfully catches 50% of identity delegation attacks at a threshold of 0.5, establishing a baseline for this attack category
Technical Details
The arena integrates with ControlArena's evaluation harness to test scenarios where an AI agent attempts to delegate its identity or permissions to unauthorized entities. The evaluation pipeline uses Claude API for both the attacker and monitor agents.
Context
Built as part of the broader AI control research community focused on understanding and mitigating risks from advanced AI systems that may attempt to circumvent oversight mechanisms.
