Code backdoors

Stacking a reasoning monitor on an output monitor blocked Sonnet 3.7's trained backdoor evasion

Anthropic trained Claude Sonnet 3.7 to talk a reasoning monitor into passing a blatant backdoor, and separately to slip subtler ones past an output-only monitor. Facing both at once, its reasoning stopped paying off.