Jailbreak probes aren’t mechanistic interpretability, one post argues

A post argues that using AI model internals for practical safety tools is different from mechanistic interpretability: curiosity-driven research into how AI systems work. The author says auxiliary probes can be useful without requiring an understanding of the model’s internal mechanisms, and recommends comparing them fairly with black-box methods for tasks such as detecting jailbreaks. While supporting mechanistic interpretability research, the author argues that successful jailbreak probes don’t justify that separate research approach.