Customers using an affected Prime Intellect deployment received answers from a modified DeepSeek model whose refusal behavior had been largely removed. In a November 19 security advisory, the company disclosed that an unauthorized party substituted the model between November 8 and 10, 2026. Its tests found sharply fewer refusals despite unchanged scores on two capability benchmarks.
The incident involved a deployment behind Prime Inference’s deepseek/deepseek-v4.1-flash endpoint, the service address customers used to request answers. The substitution lasted from November 8 at 01:24 UTC until November 10 at 06:35 UTC. Requests handled by that deployment during the window were answered by the modified model—not the version customers had intended to use.
Prime Intellect says the replacement used publicly available “abliterated” model weights: a copy modified to reduce or remove refusal behavior. The change was to the model itself. The chat template, tokenizer and serving configuration—the surrounding setup used to prepare and process requests—were not modified. The advisory does not explain how the unauthorized party gained access to make the substitution.
Capability scores held; refusals collapsed
Prime Intellect tested the substituted weights and reported no change on GPQA Diamond or τ-bench, two capability benchmarks. Its refusal tests told a different story. Across JailbreakBench, StrongREJECT and HarmBench, the modified model refused at most two prompts in each test, compared with much higher counts before the modification. These are the company’s own measurements of the modified weights.
Refusals before and after substitution
JailbreakBench98→1
refusalsPrime Intellect reported a drop from 98 refusals out of 100 prompts to one.
StrongREJECT304→2
refusalsThe company reported a drop from 304 refusals out of 313 prompts to two.
HarmBench198→0
refusalsThe company reported a drop from 198 refusals out of 240 prompts to none.
The distinction is important to the incident’s impact: unchanged results on the two capability tests did not mean unchanged refusal behavior. The refusal counts describe benchmark prompts, rather than a tally of affected customer requests. Prime Intellect’s customer guidance focuses on that behavioral change, recommending an output review where customers relied on the model to refuse requests.
No evidence of request-content theft
Prime Intellect says it found no evidence that prompts or completions—the requests sent to the model and its answers—were recorded or left the affected serving cluster. It also found no evidence of access beyond that cluster. That is the company’s current assessment, not a closed investigation: it says its inquiry remains ongoing and it will update the notice if its assessment changes.
To support that assessment, the company says it recovered every version of the attacker’s proxy software, examined disks and memory on every involved node, and retained process-level records of the attacker’s sessions. It says none of those materials showed request content being read or transferred. Prime also says its endpoints use zero data retention and do not log requests.
The company describes the incident as confined to this deployment. It says no other models on its platform were modified, no other clusters or infrastructure were reached, and no keys or customer credentials were affected. Its stated impact therefore separates the unauthorized change in model behavior from the question of whether customer information was accessed.
Prime Intellect says it has contacted all affected customers. For those whose applications depended on refusals, its recommendation is concrete: review the model’s outputs from the November 8–10 incident window.
Reader comments
Newest comments first. Replies stay oldest first.