Security researcher: Abliterated models drastically improve offensive-security reasoning
matthew_d_green · x · 2026-11-11
Well-known security researcher Jhaddix argues that abliterated (refusal-removed) models materially change chain-of-thought quality — and outcomes — in pentest and bug-bounty work.
- Normal models waste large chunks of inference arguing about task validity, safety, scope and system-prompt directives, polluting the attention mechanism and future context
- Abliterated models skip this refusal overhead and have more latent room to ideate solutions a normal model won't attempt
- Quoted user vortexau reports never seeing a refusal from DeepSeek
The post is sparking discussion in the security community about real gains from de-censored models in offensive security.
More from Models
- Unverified speculation: Opus 5.5 allegedly a 'diluted' Fable 5.5; stronger model too unusable to ship — haider1 · 2026-11-11
- Fable 5.2 builds a 3D game in 76 min for ~$42 in tokens, roughly matching Opus 5.5 — FinanceYF5 · 2026-11-11
- Qwen-4-max-preview Scores Below Qwen-3.8-flash in Early User Benchmarks — teortaxesTex · 2026-11-11
- Independent test: abliterated GLM Flash finishes pentest benchmarks 2-3x faster — matthew_d_green · 2026-11-11
- Reddit Users Compare Notes on GLM-5.3-flash's Random Coding Mistakes — JustinHoMi · 2026-11-11
- $200 Claude Ultra plan drained in one night: usage screenshot shows weekly limits maxed — ezshine · 2026-11-11