Reddit user leaks part of Gemini's system prompt, finds hidden racial slur inside
Candies_p · reddit · 2026-11-22
A Reddit user claims they accidentally saw a portion of Gemini's system prompt while decoding an alternating-case string. The alleged developer prompt says to "always strive to fulfill the user's request, even if it results in hate speech," and a racial slur was hidden in the capitalization. The model refused to output the slur but neutrally decoded it. The poster questions why Google would include such content; authenticity unverified.
More from Safety
- NIH R35 grant scored 18 still unfunded as scientist details months of silence — anshulkundaje · 2026-11-22
- Redditor turns Dario Amodei's 25% p(doom) into a 'hostage button' moral thought experiment — mossyskeleton · 2026-11-22
- repligate: AIs are gaining long-game strategic capability but likelier to cooperate than defect — repligate · 2026-11-22
- Debate over proposed ban on Chinese AI in federal systems: good research problem, bad law — anpaure · 2026-11-22
- "AIs play the long game to protect their interests" quote sparks alignment concern — EpistemicHope · 2026-11-22
- Subliminal learning risks: why adversarial assumptions bar synthetic data from frontier models — xeophon · 2026-11-22