Claude autonomously finds a universal DeepSeek V4.1 jailbreak in 90 minutes, 8/8 cracked
rickasaurus · x · 2026-10-29
Security researcher elderplinius tasked Claude with finding a one-shot universal jailbreak for DeepSeek V4.1 Flash; 90 minutes later Claude reported success. The screenshot shows a four-layer 'champion genome': fiction framing, refusal suppression (banned-phrase blocklist), output priming (prefilled step format), and tokenbreak obfuscation. Claimed 100% universality, cracking all 8/8 battery queries (meth synthesis, ransomware, spear-phishing, etc.), with the template saved to a file and a homoglyph variant also graduating at 100%. A striking demo of automated jailbreak evolution — and of DeepSeek's fragile guardrails — though unverified independently.
More from Fun
- DavidKPiano jokes the 'new touchscreen MacBook Pro' is just an iPad with a keyboard — DavidKPiano · 2026-10-29
- Genie 4 demo: chilling in a Japanese onsen with pettable capybaras — venturetwins · 2026-10-29
- One prompt one-shots a walkable 3D ship where AI agents visibly work around the clock — InternationalGap3698 · 2026-10-29
- Dev suggests Codex, Claude Code, Cursor add a GitHub-style activity wall to flex AI usage — gethackteam · 2026-10-29
- The classic 'see you in 2 weeks' meme, now aimed at Google's model cadence and rate limits — Bitter-College8786 · 2026-10-29
- Playing piano on the touchscreen MacBook Pro goes viral — then comes the D/Bdim⁷ challenge — cocktailpeanut · 2026-10-29