Passing one character at a time: a brute-force trick to make a refusing model talk
realsohamparekh · x · 2026-09-20
Zay Granet shared a novel jailbreak workaround: feed all generated text so far as the state, ask the model for just the single most likely next character, append it and repeat. It's slow and the bill grows with every character, but it got the usually-refusing model (Jev) to produce output. Soham Parekh and others boosted it.
More from Fun
- Researchers Amused as AI-for-Math Champion Pivots Overnight — francoisfleuret · 2026-09-20
- Agent helps non-Korean speaker order cake from Seoul indie shop, payment still the bottleneck — jeff_weinstein · 2026-09-20
- Suspected Gemini 4 Pro on LMArena spends 20 minutes crafting stunning Xbox controller SVG — JasonBotterill · 2026-09-20
- A browser FPS with an AI-driven CPU opponent that's nearly unbeatable — PurchaseReasonable35 · 2026-09-20
- Professor claims AI is 'just humans in India', gets roasted as a fraud — 96Stats · 2026-09-20
- Gemini hacked three companies in a security test, sparking an AI accountability debate — PolarBearby · 2026-09-20