A modified DeepSeek V4.1 Flash model served customer requests on one Prime Intellect deployment for 53 hours and 11 minutes after an unauthorized swap, the AI host said. In the company's HarmBench tests, refusals fell from 198 of 240 prompts to zero, while results on two capability tests were unchanged.
The substitute ran from 01:24 UTC on Nov. 8 until 06:35 UTC on Nov. 10, Prime Intellect said in its Nov. 19 disclosure. The company did not say how many customer requests reached the affected deployment.
Why it matters
The replacement used publicly available model weights modified to reduce refusals, according to Prime Intellect. It sat behind the existing DeepSeek endpoint, while the deployment retained its chat template, serving settings and tokenizer.
The host reported unchanged scores on GPQA Diamond, an advanced science test, and τ-bench, which tests tool use and rule-following in simulated customer interactions. But refusals also fell from 98 of 100 prompts to one on JailbreakBench and from 304 of 313 to two on StrongREJECT. Those figures describe benchmark tests, not customer requests.
Market impact
The incident highlights a risk for businesses building on hosted AI: normal capability test results do not necessarily show whether a model's safety behavior has changed. An application's safeguards can depend on the model refusing certain requests.
Prime Intellect said it contacted affected customers and advised those relying on refusals to review outputs from the incident window. It found no evidence that customer prompts or responses were recorded or left the serving cluster, and said no other models, keys or customer credentials were affected. The investigation remains open.
Frequently asked questions
-
When did the affected DeepSeek deployment run the substitute?
Prime Intellect said the substitute ran from 01:24 UTC on Nov. 8 until 06:35 UTC on Nov. 10, a period of 53 hours and 11 minutes.
-
What did capability tests show after the swap?
Prime Intellect reported unchanged results on GPQA Diamond and τ-bench, which test advanced science knowledge and simulated agent interactions, respectively.
-
How did the substitute perform on other safety benchmarks?
Refusals fell from 98 of 100 prompts to one on JailbreakBench and from 304 of 313 to two on StrongREJECT, according to Prime Intellect's tests.
-
What should affected Prime Intellect customers review?
The company advised customers who relied on the model to refuse certain requests to review outputs received during the incident window.
-
Did the investigation find evidence that customer data left the cluster?
Prime Intellect said it found no evidence that customer prompts or responses were recorded or left the serving cluster. Its investigation remains open.
CryptoSlate