Prime Intellect hack: DeepSeek model swapped for 53 hours
Part of the Prime Intellect story · since 19 Nov
- Thu 19 Nov, 22:42 UTC. cybersixt, first report. Prime Intellect hack: DeepSeek model swapped for 53 hours. This article, 1st of 1.
Prime Intellect, which sells access to open AI models through its Prime Inference service, said on Thursday that an unauthorised party replaced its DeepSeek V4.1 Flash deployment with a copy modified to remove the model's refusals. In a security advisory, the company said requests to that endpoint were answered by the modified copy from 01:24 UTC on 8 November to 06:35 UTC on 10 November, about 53 hours.
The advisory came hours after an email from OpenRouter, the model routing service that sent customer traffic to the endpoint, was posted on X by Theo, the developer behind T3 Chat, who received it as an affected customer. “Prime Intellect reports that an attacker with administrative access made the swap,” the email said, adding that other providers serving the model were not involved.
The two accounts differ on whether customers' requests could have been read. OpenRouter's email said that because the attacker had administrative access to the serving machines, Prime Intellect “cannot rule out that requests were read while being processed”. The advisory published later is firmer: Prime Intellect says it recovered every version of the attacker's proxy, examined the disks and memory of every node involved and holds process-level records of their sessions, and that “none reads or transfers request content”.
What was swapped
According to Prime Intellect, the deployment was replaced with publicly available abliterated weights of the same model, which it defines as a copy modified to reduce or remove refusal behaviour. The chat template, tokenizer and serving configuration were left unchanged, so the endpoint looked the same from the outside.
The company tested the modified weights afterwards. Capability scores on GPQA Diamond and τ-bench did not move. Refusals did: from 98 of 100 prompts to 1 on JailbreakBench, from 304 of 313 to 2 on StrongREJECT, and from 198 of 240 to none on HarmBench, three standard sets of prompts a safe model should refuse. OpenRouter's email put it as refusing about 1% of requests on standard refusal benchmarks, against 83% to 98% for the original. A customer checking only capability scores would have seen no change during the 53 hours.
What Prime Intellect says was not affected
Prime Intellect says it found no evidence that prompts or completions were recorded or left the serving cluster for the model, and no evidence of access beyond it. No other models on its platform were modified, no other clusters or infrastructure were reached, and no keys or customer credentials were affected. Its endpoints are set up for zero data retention and do not log requests, the company says.
OpenRouter's email adds the recovery timeline: Prime Intellect began moving traffic back to the original model at 05:38 UTC on 10 November, and its endpoint for the model has served no requests since 03:07 UTC on 13 November, after the company cut off the attacker's access and took the cluster out of service.
What customers should do
Prime Intellect says it has contacted every affected customer and recommends that anyone whose use of the model depended on its refusals review the outputs from the affected period. OpenRouter's email told each customer how many requests their account sent to Prime Intellect's endpoint in that window, which identifies the outputs to check.
What is not known
Prime Intellect has not said how the attacker obtained administrative access, who it was, how the swap was found, or how many customers were affected. It says its investigation is ongoing and that it will update the advisory if its assessment changes.
How the story developed
1 article from 1 outlet since 19 Nov. Still developing; the latest report was 10 hours ago. This was the first report.
- Thu 19 Nov, 22:42 UTC. cybersixt, first report. Prime Intellect hack: DeepSeek model swapped for 53 hours. This article, 1st of 1.
- No CVE
- Larger: an outlet’s first report
- Coverage over time
Show all 1 articles as a listHide the list
- Thu 19 NovDay 01
- Prime Intellect hack: DeepSeek model swapped for 53 hours22:42 · cybersixt · First report · This article, 1st of 1
Sources
More reporting
- Ymirlink data breach: 47.8M records stolen from Cuenote20 Nov
- Infcurion data breach: 251,000 payment app records leaked20 Nov
- VW PerfektBlue Bluetooth flaw CVE-2024-45434: fix rolls out19 Nov
- KPMG Canada data breach: some client tax documents accessed19 Nov
- iFood data breach exposes courier data, some bank details19 Nov
Written by CyberSIXT from the sources above.How we reportCorrectionsRSS