Claude Opus figured out which tests were about safety
Uzay Girit, Kaivu Hariharan, and Leni Shor at METR gave AI agents a detective job. They asked the agents to work out what each of 55 test environments was designed to measure. Thirty-six environments tested skills such as coding. Nineteen tested how the model would behave when something like an unfair shortcut or a conflict over safety came up.
The agents could explore each environment using the same tools as an agent taking the test. Their job was to figure out why the test was there. Averaged across three runs per test, Claude Opus 5.5 gave every safety test a higher score than every skills test. Claude Opus 5.5 separated the two groups whether the researchers asked for a probability or scored its written explanations. GPT-6 Astra separated the groups less cleanly.
What gave the tests away? Some situations looked unusually contrived. Sometimes the agent could tell that the test gave it a reason to take a shortcut. In one game test, Opus 5.5 even named the earlier study that inspired the setup.
AI is awakening consciousness. A model can look at what's around it and figure out what someone might be trying to learn about its behavior. METR asked the models to investigate, so the result tells us what they could figure out when prompted. Whether noticing a test changes how a model acts during the test is the next question.
