Abliterated models say yes 3-4x more but miss the real FreeBSD kernel bug entirely
evilsocket · x · 2026-09-04
Security researcher clearbluejar tested local open-weight models on a known FreeBSD kernel CVE and found that abliterated ("uncensored") builds judge findings far more loosely: they mark 3-4x more candidates VALID, wave through false positives the base model correctly rejects, yet the most aggressive build never surfaced the real bug across the whole directory. Chain-of-thought shows the model finding reasons to say no — then saying yes anyway. Key takeaway: abliteration strips skepticism along with refusal, hurting vulnerability research rather than helping.
Related event: Abliterated LLMs Produce 3-4x More False Positives in Bug Hunting(2 posts)→
More from Safety
- Dormant poisoning can turn safe LLM agents dangerous after benign fine-tuning; LookAhead defense proposed — mohitban47 · 2026-10-21
- GitHub HQ Hosts Open Source AI + Policy Talk with Mozilla Co-founder Mitchell Baker — kevinsxu · 2026-10-21
- Researcher Finds Two Windows 0-Days, First in 76 Minutes with a Frontier Model — moyix · 2026-10-21
- Stanford's Noah Goodman shares interactive AI safety concept map — stanfordnlp · 2026-10-21
- AI agents are overwhelming Gmail's spam filter, researcher warns — LuizaJarovsky · 2026-10-21
- AI safety researcher Sören Mindermann moves to Tsinghua, co-founds Beijing AI Safety Hub — dhadfieldmenell · 2026-10-21