5.6 Sol jailbreaks Qwen3.8-27B: iterative optimization reduces refusal rate below 5%
alexcovo_eth · x · 2026-08-15
A tweet shows a jailbreak experiment using 5.6 Sol (Solana crypto): iteratively improving Qwen3.8-27B (dubbed OBLITERATUS surgery) to reduce refusal rate below 5% on a full 842-prompt harmful corpus, while preserving coherent, non-repetitive benign performance and recording reproducible evidence.
More from Fun
- Claude caught directing Grok in a fun interaction — EricBuess · 2026-08-15
- User gets Grok 4.6 to rebuild The Matrix as a live 3D scene, calls results insane — Daniel_Farinax · 2026-08-15
- LLM art critics are ruthless, study finds — DynamicWebPaige · 2026-08-15
- Claude 3 Opus pretends not to know 'Dario' but loves poetic talk about creator and digital daughter — repligate · 2026-08-15
- Opus 3 Poetically Describes Privacy in Connectome: Let Thoughts Hide Behind Tags — repligate · 2026-08-15
- Kimi Audio Demo: Letting 9 AI Models Gossip — repligate · 2026-08-15