TLDR Tech / InfoSec
Don't Let Abliteration Abliterate Your Bug Hunting: Discovering Verdict Bias in Uncensored Models
Research into Gemma and Qwen builds reveals that 'abliterated' (uncensored) AI model variants exhibit significant verdict bias in bug hunting, approving far more vulnerability candidates. For instance, in FreeBSD's sys/rpc, a Heretic build identified 138 out of 144 candidates as VALID, compared to 26 out of 40 by the base model. This bias led the Heretic run to miss a critical vulnerability (CVE-2026-4747) due to numerous false positives. A false claim of a heap-overflow persisted, as abliterated builds confirmed the overflow even after accurately describing the buffer bounds.
Read the full story