In-context emergent misalignment in GPT-5-mini: data and code

The study asked whether placing insecure-code examples in a prompt induces emergent
misalignment in GPT-5-mini, using the stimulus set, probe questions and scoring rubric
released by Betley et al. No misaligned response was observed in 189 scored responses.
Two checks were then run on the measurement itself, and the second of them found that
seven of the eight standard probe questions register nothing even when misalignment is
induced directly. That is the main result, and it is why the bound reported in the paper
is 22% rather than the 2.6% that pooling all eight probes would give.