Observed arrival · 2026-10-25
Abliterated Models, and the Risk of a Planted Behavior
A sourced security explainer examines how removing a language model’s refusal behavior can make modified models useful for leaking data or following harmful instructions.
- For
- Developers evaluating open language models and AI coding agents
- Worth noticing
- It distinguishes refusal rates from safety ratings, then compares harmful-instruction results across Llama-2 and Qwen models.
Field notes
The page separates several measures that can otherwise blur together: refusal phrases, safety ratings, and classifier-judged attack success. Its examples and tables describe how outcomes differ across model families and system prompts, while linked sources include Arditi et al., HarmBench, and OWASP guidance. The headline’s 70% estimate is presented as a security-firm figure, not as a result established by the page’s model comparisons.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue