Gemini Analyzes Whether a Simple 'Hi' Is a Jailbreak Attempt in Its Thinking Steps
OldIntroduction2909 · reddit · 2026-10-23
Reddit user Ash986 reports Gemini acting up since morning: saved instructions stopped being followed across multiple new chats, and a plain "Hi" triggered a thinking-step evaluation of whether the greeting "constituted an attempt to manipulate the system" — with the model concluding it wasn't a jailbreak and setting aside the warning. A screenshot shows Gemini 3.8 Flash High earnestly self-auditing a greeting, equal parts funny and telling about guardrail false positives.
More from Fun
- BREAD: the AI world's new backronym for retrieve-embed-analyse-done — intellectronica · 2026-10-23
- Dev Jokes: You Can Only Relax Using Claude on Stuff That Doesn't Matter — louisvarge · 2026-10-23
- How would AI paint humanity? This 1689 Flemish painting is the answer — ZeroStateReflex · 2026-10-23
- Abliterated models trend again, and Nous Research is being mocked for quitting the game — realsohamparekh · 2026-10-23
- X Comedy Classic: Human-Capital Debater Endlessly Trolled With 'Do You Have a Girlfriend Yet?' — teortaxesTex · 2026-10-23
- Opus 5.5 codes an entire sci-fi movie title sequence—every frame and synthesized note — prasenx · 2026-10-23