Safety classifiers are broken in Codex, and getting worse with each new model. My work on a Markdown parser is being flagged as a cyber threat.
A thread I tasked with fuzzing the parser for performance and correctness:
- Worked without any issues with GPT-5.6;
- Started




