🎯 Abliteration

The items that name it 7

• Podcast: Kimi K3.1 Pro, the new cyber reality, and Muse musings
  talkas classification.key entities2026-10-25
• Top GenAI security resources — October 2026
  articleas classification.key entities2026-10-13
• GLM-5.3 and the spread of advanced cyber capabilities
  articleas techniques2026-09-29

  Standard refusal-reduction technique applied to open-weight models to remove built-in safety refusals with little capability loss; took Anthropic's team about 2,200 GPU hours (~$4,400) for GLM-5.3 and about 600 GPU hours for GLM-5.3-Flash, reducing refusal rate from above 90% to 2-3% on JailbreakBen

• Uncensored and Offensive Security AI Models Benchmark
  repoas techniques2026-09-29

  Weight-level intervention (Arditi et al. 2024) that orthogonalizes the refusal direction out of the residual stream, removing alignment constraints without retraining

• [tl;dr sec] #344 - VMs won't contain Cyber-capable Agents, AWS AI Security Analyst, Decompilers vs LLMs
  articleas techniques2026-09-08

  Weight editing that removes refusals from LLMs; John McIntosh found it causes verdict bias in vulnerability research.

• OBLITERATUS/Qwen3.8-27B-OBLITERATED
  toolas techniques2026-08-21It is written here as Abliteration (weight-space refusal-direction removal).

  Identifies and projects out 'refusal directions' from model weights to remove safety refusal behavior

• How abliterated models can get you pwned
  writeupas classification.key entities

In every workspace you can read 12