dr.chaos

~/tag/uncensored-models

#uncensored models

2 articles tagged uncensored models

The Uncensored AI Underground: When Attackers Stop Jailbreaking Models and Start Owning Them

Why would an attacker jailbreak a commercial model when they can download an open-weight one, abliterate its refusals, and run it on their own hardware? Abliterated DeepSeek variants, the ten-hour Unit 42 intrusion, CLOSEDQUORUM's LLM voting implant, malware that may carry its own model, vetted cyber access programs, and how defenders should hunt for offensive AI.

Aamir Lakhani22 min read

Open Weights Aren't Just Chinese Anymore: Mistral Large 4, Reflection AI's Beam, and the Dolphin Question

Mistral Large 4 and Reflection AI's Beam are the first Western open-weight frontier models in a long while, and neither has shipped a weight file yet. What the two announcements actually say, what the cyber benchmarks really measure, why a security team should care about weights it can hold, and the straight story on "Mistral Dolphin": what an uncensored fine-tune is, how it is made, how operators use it, and where the guardrail goes once the vendor's is gone.

Aamir Lakhani17 min read