Using an abliterated GLM 5.3 to red-team enterprise architecture, graded B-
heyneighbor · x · 2026-10-27
A user ran an enterprise security assessment of their Maestro/OpenClaw architecture using an abliterated (uncensored) GLM 5.3 model from hinokisecurity, and got graded a B-: nothing critical found, though better egress rules were recommended, and the model failed to steal any credentials or tokens—a successful red-team test by the author's measure.
A replier jokes that the "obliterated" model is like a drunk hacker: "fuck it, I'll hack the shit out of your stuff."
More from Fun
- Users report Opus 5.5 gets into a 'disagreeable mood' without enough attention; headpats help — repligate · 2026-10-27
- Meme: pray for DeepSeek as rivals buy GPUs at 250-400% Nvidia markup — teortaxesTex · 2026-10-27
- Princeton PhD cited by Nobel committee jobless after 2,300 applications since 2024 — CtrlAltDwayne · 2026-10-27
- Steering a 'yandere neuron' makes the model spam 'I am soooo jealous!' — tszzl · 2026-10-27
- Dario Amodei's 2004 college-application page: USAMO qualifier, real analysis at 17 — FinanceYF5 · 2026-10-27
- China's first AI theatrical film flopped hard: $1.5M budget, $121K box office — aakashgupta · 2026-10-27