Models That Will Not Refuse: Dark and Abliterated AI
An annex to CB-288 on AI models with refusal safeguards removed: wrapped jailbreaks, criminal fine-tunes, abliterated open models, hosted unrestricted services and poisoned models. It explains why refusal is cheap to strip, what unrestricted models change for pretexting, research and tooling aimed at a principal, and what they do not change. It closes with questions for households, family offices and vendors, and the controls that hold regardless of model behavior. For security leads and family offices.
Unlock this Command Brief.
Get this assessment, the full subscription collection, and new intelligence as it is published.
For your first year, then $100/month. Or $750 for the first year, then $1,000 annually.
- The CORGI Morning Brief, each morning
- This brief and the subscription collection, 90+ briefs
- New Command Briefs as they are released
- Principal and Analyst briefs, as PDF
Secure checkout through Substack. Substack asks for your email first.
How it happens
AI-composed claims
Generative systems compose statements about people, which can attach another person's history and be repeated until they look confirmed.
More on infer and profile →- Mechanism · Deceive and engineerPretext assembly. An approach is built from small true details gathered across sources, so it reads as routine.