For about 53 hours this month, requests to Prime Intellect’s DeepSeek V4.1 Flash endpoint were answered by a different model. An intruder had swapped it for a copy with its safety refusals stripped out, and its scores on standard quality tests stayed the same. Prime Intellect’s security notice puts the window at 01:24 UTC on 8 November to 06:35 UTC on 10 November.
Prime Intellect is a startup that sells access to open AI models through its Prime Inference service, CyberSIXT reports. Some customers reached it through OpenRouter, a service that sends each API call to one of several providers hosting the model.
What the swap changed
The replacement was a publicly available “abliterated” copy of the same model, which Prime Intellect describes as modified to reduce or remove refusals. The chat template, tokenizer and serving settings were untouched, so the endpoint looked identical from outside.
In Prime Intellect’s own tests, scores on GPQA Diamond and τ-bench, two capability tests, did not move. Refusals collapsed. On JailbreakBench they fell from 98 of 100 prompts to 1, on StrongREJECT from 304 of 313 to 2, and on HarmBench from 198 of 240 to none, all three being sets of requests a safe model should turn down. One quality score says little about a model, as our look at what AI leaderboards measure explains.
Who heard first
Customers heard from OpenRouter before Prime Intellect said anything in public. On 19 November, Theo Browne, the developer behind T3 Chat, posted OpenRouter’s email to affected users on X. It said “an attacker with administrative access made the swap” and that other providers serving the model were not involved. It also gave each customer a request count for the window; Theo’s account had sent 3.
Prime Intellect published its notice later that day, Times of AI reports. OpenRouter’s email says traffic began moving back to the original model at 05:38 UTC on 10 November. The endpoint has served no requests since 03:07 UTC on 13 November, and Prime Intellect says it cut off the attacker and took the cluster out of service.
What Prime Intellect says was not touched
The company says it found no evidence that prompts or answers were recorded or left the cluster, and no access beyond it. No other models changed, and no keys or customer credentials were affected. It says it recovered every version of the attacker’s proxy and examined every machine involved.
OpenRouter’s email was more cautious: because the attacker had admin access, Prime Intellect “cannot rule out that requests were read while being processed.” All of this comes from the company itself, and we found no independent forensic report. How the attacker got in, who it was and how many customers were hit are still unknown.
A second warning the same day
Also on 19 November, Dragos research chief HD Moore published a study in which his team sent about two million prompts laced with decoy links and credentials to AI services. Dragos says it saw evidence that data from at least three AI providers was exposed. It names no providers and does not mention Prime Intellect.
Both stories point at the middle layer between you and the model. That matters for anyone sending company data to AI tools, which our checklist for customer data in AI tools covers.
What to do this week
If you called DeepSeek V4.1 Flash through OpenRouter between 8 and 10 November, find OpenRouter’s email with your request count. Review those outputs if your product relied on the model refusing harmful requests. Then keep a few fixed “should refuse” test prompts and run them against every hosted model you use, next to your quality checks.
Sources
Get the weekly email
The week's most useful AI news, a tutorial and an interview.

