An unauthorized party replaced an AI model hosted by Prime Intellect with a copy altered to refuse fewer requests, or none at all. The copy answered customers for 53 hours, from November 8 to 10, 2026. When the company tested it afterward, it scored the same as the original on capability tests but almost never refused a harmful request. Prime Intellect, which trains and hosts freely downloadable AI models, disclosed the incident in a security advisory on November 19.
The model in question is DeepSeek V4.1 Flash, a language model of the kind that powers chatbots. The Chinese company DeepSeek released it as an open-weight model, the category at issue in this case. Its weights, the files that hold what a model learned in training, are published for anyone to download and run on their own machines.
The copy that replaced it was public too. It was an “abliterated” version, one whose refusal mechanism has been removed without retraining the model. Dozens of copies like it are available on Hugging Face, the main platform for sharing AI models.
A model swapped for more than two days #
Prime Intellect sells access to open model families such as GLM, DeepSeek and Kimi through a service called Prime Inference. Customers pay according to the amount of text the models process. Their software connects to the service the same way it would connect to OpenAI’s API, the interface that programs use to send requests to OpenAI’s models. The advisory, numbered PI-SA-2026-001, covers the deepseek/deepseek-v4.1-flash endpoint, the address customers use to call that model.
According to the advisory, the intruder replaced “a deployment” behind that endpoint with a modified, publicly available copy of the model. A deployment is one installation of the model on the company’s servers. The swap lasted from 1:24 a.m. UTC (Coordinated Universal Time) on November 8 to 6:35 a.m. UTC on November 10. For those 53 hours and 11 minutes, the copy answered the requests handled by that installation.
For customers, the model name and the format of the exchanges stayed the same. The advisory says the conversation format and the server settings were not changed. Neither was the tokenizer, the tool that splits text into tokens, the small chunks of text a model reads and writes. Prime Intellect says it contacted every affected customer. It advised those who relied on the model’s refusals to review the answers they received during the swap.
One public copy among dozens #
DeepSeek posted the model on Hugging Face on September 10, 2026, under the MIT license, which allows commercial use and modification. Its model card, the description page published with it, says the model’s main network holds 552 billion parameters. Parameters are the numerical values set during training. The model can read up to one million tokens at once.
The technique behind these copies, abliteration, builds on work by researcher Andy Arditi and his co-authors. In the chat models they studied, refusal depends on a single pattern in a model’s internal computations. They call this pattern a “direction.” Neutralize that direction and the model stops refusing, without any retraining.
The authors first described the finding in a blog post. On June 17, 2024, they posted a paper on arXiv, an open online archive of scientific papers. Four days earlier, on June 13, researcher Maxime Labonne had described the technique on Hugging Face under the name “abliteration,” citing the blog post.
A Hugging Face search on November 22, 2026, listed 67 public repositories, or download pages, whose names contain both “V4.1-Flash” and “abliterated.” Most of them offer a modified copy for download. Seventeen are built on the September model, the one named in the advisory. The other 50 are built on a newer version DeepSeek released on October 17.
The first of these repositories was created at 10:59 a.m. UTC on September 10, 2026, less than nine hours after DeepSeek posted the model. Prime Intellect’s advisory does not say which public copy was loaded or where it came from.
Same capability scores, almost no refusals #
When Prime Intellect tested the modified weights after the incident, the copy matched the original on two capability tests. The first, GPQA Diamond, is a set of advanced science questions. The second, τ-bench, measures whether a model can use software tools in simulated conversations with customers.
On three safety tests, the copy’s refusals all but disappeared. The company used JailbreakBench, StrongREJECT and HarmBench, which send a model harmful requests to check that it refuses them. On JailbreakBench, the original refused 98 of 100 requests, and the copy refused one. On StrongREJECT, refusals fell from 304 of 313 requests to two; on HarmBench, they fell from 198 of 240 to none.
In its November 19 report, the AI news site Superpower Daily wrote that unchanged capability scores did not mean the model behaved the same way. That means a check limited to those two capability tests would have flagged nothing.
No sign of a leak, the company says #
Prime Intellect says it found no evidence that requests or responses were recorded. It also says it saw no sign that they left the group of servers running the model, or that the intruder got access beyond that group. According to the advisory, no other model on the platform was modified, and no access keys or customer credentials were affected.
The company adds that its endpoints are set up for “zero data retention,” so they keep none of the data they process, and that they do not log any requests. It says it recovered every version of the intruder’s proxy, a program that relays requests, and examined the disks and memory of every machine involved. It also holds detailed, process-level records of the intruder’s sessions. According to the company, none of the proxy versions reads or transfers request content.
This incident is different from the October breach at Fireworks AI, another company that hosts AI models. In November, Notion, the maker of the note-taking app, confirmed that conversations with its Notion AI assistant had leaked in that breach. At Prime Intellect, the intruder changed the model that answered customers, and the company says it found no data leaving its servers.
Gaps in the advisory #
The advisory does not explain how the intruder got the access needed to change the model. Superpower Daily calls this the main gap in what is known so far. The advisory also leaves out how many requests the copy served, how many customers were affected and how the swap was detected. The investigation is ongoing, and Prime Intellect says it will update the advisory if its assessment changes.
What it means #
The incident exposes a weakness specific to open models accessed through a hosting company and billed by use. A closed model’s developer keeps its weights private and shares them only with a few partners it chooses. The weights of an open model can be loaded by any host, and modified copies circulate freely. Customers pick a model name, but they never see which weight files are loaded on the server.
The number of middlemen adds to the risk. OpenRouter, a service that passes requests on to different hosts, listed 37 endpoints run by 28 hosts for DeepSeek V4.1 Flash alone on November 22, 2026. A single model name can cover dozens of installations whose files the customer never sees. The chain keeps getting longer: since November 16, Hugging Face has let OpenRouter route its customers’ requests to language models.
The test results Prime Intellect published show that the copy behaved differently from the original. The assurances about customer data rest on the company’s internal investigation, and the advisory mentions no outside audit.
A separate risk concerns applications that relied on the model’s refusals. For 53 hours, they may have passed along answers the original model would have refused.
Running the model in-house offers a check that a host’s customers do not get. Hugging Face publishes a digital fingerprint, called a SHA-256 hash, for each of the 48 weight files of DeepSeek V4.1 Flash. The hash is a string of characters computed from the file, and it lets anyone confirm that a downloaded file is DeepSeek’s own. That check does not protect against an intruder who takes control of the server.
The software that runs a model has flaws of its own. vLLM, an open-source program that runs models on servers, published 58 security advisories between September 1 and November 16. Prime Intellect’s advisory does not say which software the company uses.
The advisory still leaves two questions unanswered: how the intruder got in and how many customers received answers from the copy. It is also unknown whether other hosts of open models check the weights they serve, and whether they tell customers what they find.
Sources #
- [Notice regarding deepseek/deepseek-v4.1-flash, 8 to 10 November 2026](https://www.primeintellect.ai/security/advisories/deepseek-v4-1-flash-model-substitution) , Prime Intellect, November 19, 2026
- [Prime Inference](https://www.primeintellect.ai/inference) , Prime Intellect, accessed November 22, 2026
- [DeepSeek-V4.1-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) , DeepSeek on Hugging Face, September 10, 2026
- [Search “DeepSeek-V4.1-Flash-abliterated”](https://huggingface.co/models?search=DeepSeek-V4.1-Flash-abliterated) , Hugging Face, accessed November 22, 2026
- Refusal in Language Models Is Mediated by a Single Direction , Andy Arditi et al., arXiv, June 17, 2024
- Uncensor any LLM with abliteration , Maxime Labonne, Hugging Face, June 13, 2024
- Prime Intellect discloses unauthorized DeepSeek model swap that sharply reduced refusals , Superpower Daily, November 19, 2026
- DeepSeek: DeepSeek V4.1 Flash, providers , OpenRouter, accessed November 22, 2026