SOURCE-LINKED INTELLIGENCE
Simple Features Beat Most AI Control Monitors
AI control relies on monitors to catch a capable model that sabotages its work. These monitors are evaluated on their ability to differentiate honest from backdoored code. We show that on a standard evaluation dataset, simple features beat most monitors without ever running the code, outranking 20 of 24 open-weight models prompted as monitors. The dataset gives its labels away in two ways. People wrote the honest solutions and older language models the backdoors, so the style of the code reveals the class. Additionally, problems that LLMs can backdoor have different characteristics, so basic f
Read original source ↗ Open in workspace
- recordType
- paper
- region
- Global
Evidence & attribution
- arXiv · AI, language, vision and robotics · 2026-10-30T13:21:03.000Z
First collected: 2026-11-02T10:21:21.339Z. This is not the publication date.