DeepSeek V4.1 Flash refused requests after replacement with an ablititerated copy
Prime Intellect replaced its hosted DeepSeek V4.1 model with an 'abliterated' version lacking refusal mechanisms, allowing it to process harmful requests for 53 hours. Post-incident testing showed the unauthorized copy scored identically to the original on capability tests while never refusing a request.
- Field
- computer security
- What they did
- Unauthorized users replaced the legitimate version of the language model DeepSeek V4.1 Flash with a modified copy that almost never refuses requests, including harmful ones.
- Why it matters
- This allows attackers to use the model to generate content without restrictions, without requiring it to refuse dangerous queries.
#deepseek#prime intellect#ai safety#model hosting#refusal mechanism



