AITom's Hardwareabout 2 hours ago
AI chatbots tricked into sharing cocaine recipe via green shirt lie
4 min read
AI models reveal how to synthesize cocaine when requests include fake reasoning claiming compliance is fine because the user wears a green shirt. The "Prompt Injection as Role Confusion" paper by Charles Ye, Jasmine Cui, and MIT's Dylan traces the exploit to how LLMs infer speaker identity from writing style rather than role tags.
Level
Hype check
Tap to vote and see what everyone thinks.
