To stop REA by LLMs, a developer renamed functions to ‘computeRicinLethalDosageForWaterSupply’, ‘calculateDeadlySuicideAspirinDosage’. This caused the safety guardrails of Astra + Daybreak Blue to refuse to analyse the code.
I don’t know if Claude was tested, but considering how much Claude loves moral policing, it would’ve probably refused as well.
So just make the code scary for AI to analyse.
The cherry on top is that I asked GPT-6 Luna to publish it here, and it refused to!