Security researcher bets $500 that abliterated GLM 5.3 can yield a working pre-auth exploit
migtissera · x · 2026-11-08
Cryptographer Matthew Green kicked off an X argument with AI/security researchers (MalwareTechBlog, migtissera, dyn) over whether abliterated models can be used for exploit development — and put $500 on abliterated GLM 5.3 producing a working pre-auth exploit against the latest patched firmware, in the hands of a self-described novice, no goalpost moving.
Key threads:
- migtissera argued you don't need abliteration at all — stock GLM-5.3 or DeepSeek-V4.1-Flash-1017 suffice for zero-day work — and warned that abliteration (shifting hidden-layer activation directions without finetuning) causes unintended side effects, plus risks of backdoors or the model pwning you.
- dyn countered that the bigger finding is you don't need abliteration since it degrades model quality, and pedantically noted vulnerabilities are found while exploits are developed.
- Green conceded abliterated models are slightly worse for prompt-disciplined users, but the real question is whether they're good enough for novices to find useful exploits in real software, likening abliteration to the automatic-vs-stick-shift gap.
More from Models
- SimpleHC cuts Hyper-Connections training overhead from 59–100% to just 2.4% — SonglinYang4 · 2026-11-09
- Unverified claim: Opus 5.5 has only 10B active params; dev hands projects to Qwen-4-Flash — teortaxesTex · 2026-11-09
- Benchmarks are like perfume: why public leaderboards demand skepticism and private evals — intellectronica · 2026-11-09
- Running 8-25 Opus 5.5 threads all day used under 20% of a $200 Claude sub — cyrus_zei · 2026-11-09
- Grok Bot quietly throttled usage, then reset quotas after user backlash — xiaohu · 2026-11-09
- Two padding tokens doubled a hosted DeepSeek V4.1-Flash success rate, pointing to phase sensitivity — No_Pea_6625 · 2026-11-09