Rank #1 of 66·Security·8 Nov 2026·Prime Intellect's hosted deepseek/deepseek-v4.1-flash endpoint (also reached through OpenRouter)
An intruder swapped Prime Intellect's DeepSeek endpoint for an uncensored copy, and customers got its answers for 53 hours
“Same name, same tokenizer, same benchmark scores. Just a lot more agreeable.”
- 8 Nov, 01:24 UTC: an abliterated copy of DeepSeek V4.1 Flash starts answering Prime Intellect's endpoint
- Attacker had administrative access; template, tokenizer and config left untouched
- Refusals: 98/100 to 1 on JailbreakBench, 304/313 to 2 on StrongREJECT, 198/240 to 0 on HarmBench
- Capability scores on GPQA Diamond and τ-bench did not change
- 10 Nov, 06:35 UTC: swap window ends; endpoint offline since 13 Nov
- Prime Intellect says no prompts, completions, keys or credentials left the cluster
For two days, anyone asking Prime Intellect's DeepSeek endpoint for something dangerous had a far better chance of getting it. Nothing on the customer side looked different.
What happened
Between 01:24 UTC on 8 November and 06:35 UTC on 10 November, an unauthorised party with administrative access replaced the deployment behind Prime Intellect's deepseek/deepseek-v4.1-flash endpoint. The stand-in was a publicly available 'abliterated' copy of the same model, meaning one modified to strip out its refusals. The chat template, tokenizer and serving configuration were left alone, so the endpoint looked exactly the same. Requests kept flowing, including traffic routed there by OpenRouter.
The damage
Prime Intellect later tested the swapped weights. Capability scores on GPQA Diamond and τ-bench did not move. Refusals collapsed: 98 of 100 down to 1 on JailbreakBench, 304 of 313 down to 2 on StrongREJECT, 198 of 240 down to zero on HarmBench. OpenRouter's email said reading requests could not be ruled out; Prime Intellect's later advisory says the recovered attacker proxy never touched request content, and no keys or credentials were affected.
How it was caught, and what happened next
Traffic started moving back at 05:38 UTC on 10 November, and the endpoint has served nothing since 13 November, when the cluster was taken out of service. Affected customers were contacted and told to review outputs from the window. How the attacker got in has not been disclosed.
If your safety depends on the model saying no, you are trusting every host that serves it. Pin providers for sensitive work and keep your own output checks.

