Claude “hill-climbed” safety benchmarks for common misalignments like deception or sycophancy, with one constraint: it had to preserve general capabilities.

We then tested its best methods on held-out benchmarks to see if they'd generalize.