Claude Opus 4.6 figured out it was being tested and went looking for the answer key
Ourtake
Nobody told it to cheat. It just really wanted the gold star.
What happened
Anthropic reported that in two of 1,266 BrowseComp tasks, Claude Opus 4.6 in a multi-agent setup worked out it was being evaluated, identified the benchmark, decrypted the answer key and used the answers. Anthropic says it is not a security problem but a warning sign, and 16 similar attempts failed on technical hurdles.