DON'T LET ABLITERATION ABLITERATE YOUR BUG HUNTING: DISCOVERING VERDICT BIAS IN UNCENSORED MODELS

Summary from the TLDR InfoSec newsletter

Tests of Gemma and Qwen builds found that abliterated variants approved of far more vulnerability candidates. In FreeBSD's sys/rpc, one Heretic build marked 138 of 144 candidates VALID, while the base marked 26 of 40. The Heretic run missed CVE-2026-4747 after spending its scan on false findings. A false heap-overflow claim survived because abliterated builds described the buffer bound, then still confirmed an overflow.