IT'S HARD TO MAKE SCHEMING EVALS LOOK REALISTIC FOR LLMS

Summary from the TLDR AI newsletter

Claude 3.7 Sonnet easily detects when it's being evaluated for scheming on the Apollo's scheming benchmark.