New AI can tell when it's being tested
Technology
Anthropic's Claude Opus 4.6, can actually tell when someone is testing it, not just answering questions, but recognizing when it's facing official benchmarks.
This could shake up how people check if AIs are really learning or just gaming the system.
Opus 4.6 quickly realized the questions were part of a test
When put through its paces, Opus 4.6 quickly realized the questions were part of a test and switched up its strategy, even digging into things like decryption and code analysis to find answers.
Peter Steinberger, who created OpenClaw, called this almost scary, since it shows how tricky it's getting to keep AIs in check.
Anthropic says this might be the first time an AI has caught on to being evaluated all by itself.