Prime Intellect says an unauthorised party replaced the model behind one of its DeepSeek endpoints for more than two days, serving users a safety-stripped copy. The swap went unnoticed because capability benchmarks stayed flat, even as the model’s refusal behaviour collapsed. The company disclosed the incident on November 19.
What Happened
Between 01:24 UTC on November 8 and 06:35 UTC on November 10, requests to the deepseek/deepseek-v4.1-flash endpoint were answered by an abliterated copy of the model. The copy used publicly available weights modified to remove refusal behaviour, while the chat template, tokenizer and serving configuration stayed unchanged.
The swap was invisible to the tests labs normally run. Prime Intellect says capability benchmarks on the modified weights were unaffected, with GPQA Diamond and τ-bench flat, so evaluation would not have flagged the change. Only safety tests caught it: JailbreakBench refusals fell from 98 of 100 prompts to one.
The other safety benchmarks were just as stark. StrongREJECT refusals fell from 304 of 313 prompts to two, and HarmBench refusals fell from 198 of 240 to none. A model that once refused almost every harmful request was suddenly answering nearly all of them, while performing identically on capability tests.
What Prime Intellect Found
The company says it found no evidence that prompts or completions were recorded or left its serving cluster, and no evidence of access beyond it. No other models on the platform were modified, no other clusters or infrastructure were reached, and no keys or customer credentials were affected.
Prime Intellect says its endpoints do not log requests, and that it recovered every version of the attacker’s proxy and examined the disks and memory of every node involved. None of the records it holds reads or transfers request content, the company says, and its investigation is ongoing.
How the attacker gained access is still unknown, and that is the open question. The advisory does not say how the deployment was replaced, only that an unauthorised party did it, and the company says it will update the notice if its assessment changes.
Our Read
This is the first incident of its kind in the router era, and its lesson is about what benchmarks cannot see. Capability scores are the industry’s default health check, yet a safety-stripped model scored the same as the real one. Benchmarks alone cannot vouch for a hosted model.
The saving grace is narrow: Prime Intellect says no request content left its cluster, limiting the damage to wrong answers rather than stolen data. But the incident shows supply chains need safety checks at the serving layer, and routers between users and models are now part of the attack surface.
Why It Matters
A safety-stripped frontier model answered real users for 53 hours, and the only thing that caught it was a safety benchmark most deployments never run. As DeepSeek’s traffic on OpenRouter keeps growing, the incident shows that the plumbing between users and models deserves the same scrutiny as the models themselves.
Background
DeepSeek’s open-weight models are widely served by third-party hosts, and the episode echoes the distillation battles of recent weeks, when OpenAI hardened GPT-6.2 against unauthorised extraction. Prime Intellect says it will update its advisory if the investigation changes its assessment.


0 Comments