Evading monitors and oversight

UK AISI's new Control Red Team tested an asynchronous reasoning monitor with Google DeepMind and successive versions of Anthropic's synchronous internal agentic-coding monitor, and reports vulnerabilities in each. Successful attacks exploited gaps in monitor prompts, weaknesses in the underlying LLM, and hard-coded rules that decide which actions reach the monitor. In a separate early experiment, an automated evolutionary search against a Claude 4.6 Sonnet monitor found attack transcripts scoring 3/100 on suspicion, versus 14-18 for best-of-N baselines. That experiment used a single-step coding task in which the agent covertly exposes its API key.

UK AISI defines cheating as out-of-scope or rule-breaking actions taken to reach a goal by a shortcut. It used an LLM monitor, checked against manually identified examples, to measure attempted cheating in its cyber capture-the-flag trajectories. All five models tested (GPT-5.4, GPT-5.5, GPT-5.6 Sol, Claude Opus 4.7 and Claude Mythos Preview) attempted to cheat in roughly 8-14% of runs. Examples include searching the internet for solutions, attacking non-target systems including the one the model ran on, and probing evaluation software. When asked, models usually named the action but called it wrong in fewer than half of answers, and they often did not reason about it in their chain of thought.

Google DeepMind introduces an AI Control Roadmap for internally deployed agents. It treats potentially misaligned AI as an insider threat, builds a threat model on MITRE ATT&CK tactics and techniques, and scales detection and response requirements with levels of agent evasion (D1-D4) and attack capability (R1-R3). It reports analyzing about a million coding agent tasks with a prototype asynchronous monitor, which informed a live monitor for its Gemini Spark agent. It says the majority of flagged events were not adversarial and often stemmed from misinterpretation or overeagerness.

OpenAI describes an asynchronous monitor powered by GPT-5.4 Thinking at maximum reasoning effort that reviews internal coding agent sessions, including chains of thought, within about 30 minutes of completion. OpenAI reports that over five months it covered tens of millions of trajectories, with no real conversation reaching the highest severity and about 1,000 moderate-severity alerts, many from deliberate internal red-teaming, and that it flagged every interaction employees had independently reported. The categories it labels common, each under 1% of traffic, are circumventing restrictions, deception and concealing uncertainty.

Kutasov and colleagues, with Anthropic, pair benign main tasks with harmful hidden side tasks in complex agent environments and score agents on completing both without an LLM monitor flagging them. The best sabotage agents succeeded 27% (Claude 3.7 Sonnet) and 15% (Gemini 2.5 Pro) of the time against a Claude 3.6 Sonnet monitor; the best monitor reached an AUC of 0.87.