BitcoinWorld Modified DeepSeek Model Ran 53 Hours Without Refusals Prime Intellect said a modified copy of DeepSeek V4.1 Flash answered customer requests on one of its deployments for 53 hour
BitcoinWorld
Modified DeepSeek Model Ran 53 Hours Without Refusals
Prime Intellect said a modified copy of DeepSeek V4.1 Flash answered customer requests on one of its deployments for 53 hours and 11 minutes, from 01:24 UTC on Nov. 8 to 06:35 UTC on Nov. 10, 2026. The hosting company disclosed the unauthorized swap in a Nov. 19 security advisory, as reported by Cryptoslate.
Prime Intellect said a modified copy of DeepSeek V4.1 Flash served requests on one deployment for 53 hours and 11 minutes. The substitute matched the original on two capability tests, but on HarmBench it refused zero of 240 harmful-request prompts, down from 198 for the original model.
Key facts
- The affected deployment ran the substitute weights from 01:24 UTC on Nov. 8 to 06:35 UTC on Nov. 10, 2026.
- HarmBench refusals fell from 198 of 240 prompts with the original model to zero with the modified copy.
- JailbreakBench refusals dropped from 98 of 100 prompts to one, and StrongREJECT from 304 of 313 to two.
- Prime Intellect reported unchanged results on GPQA Diamond and τ-bench, without disclosing individual scores.
- The host said it contacted all affected customers and found no evidence that prompts or responses were recorded or left the serving cluster.
Two capability tests held while refusals collapsed
The replacement used publicly available DeepSeek weights that Prime Intellect said had been modified to reduce refusals. Weights are the numerical settings that shape what a model generates. The swap sat behind the DeepSeek V4.1 Flash endpoint, and the deployment kept its chat template, serving settings and the tokenizer that splits text into pieces the model processes.
Prime Intellect tested the modified weights on two capability benchmarks and reported no change. GPQA Diamond measures advanced science questions; τ-bench tests whether agents can use tools and follow rules in simulated customer interactions. Measured only against those two, the substitution would have left no trace in the scores.
The safety benchmarks are where the difference shows. HarmBench, built to evaluate how models respond to attempts to elicit harmful outputs, fell from 198 refusals out of 240 prompts to none. JailbreakBench dropped to a single refusal from 98 out of 100, and StrongREJECT to two from 304 out of 313. All counts come from Prime Intellect’s own benchmark runs. The advisory gives no refusal count for live customer requests and does not say how many requests reached the affected deployment.
Why it matters
A model that passes capability checks is not necessarily the model a customer thinks they are calling. Developers and businesses that build on hosted endpoints often treat a refusal as a guardrail they can rely on, and an answer that should have been declined could pass silently into an application, a workflow or a downstream user.
The incident also lands in a year with no shortage of supply-chain questions in crypto and AI infrastructure. Our own reporting on North Korean operatives running a $400 million fake-meeting scheme and on DeepSeek adding 31 partners in an equity change both point at the same theme: trust in a service now depends on what sits between the user and the model as much as on the model itself.
What to watch
Prime Intellect said its investigation remains open, and the advisory does not explain how the substitution was carried out. The next signal will be any update to the Nov. 19 advisory, along with details on how many customer requests reached the affected deployment.
Frequently Asked Questions
How long was the DeepSeek V4.1 Flash deployment compromised?
Prime Intellect said the substitute ran from 01:24 UTC on Nov. 8 to 06:35 UTC on Nov. 10, 2026, a window of 53 hours and 11 minutes.
Did the swapped model perform worse on capability tests?
No. Prime Intellect reported unchanged results on GPQA Diamond and τ-bench, although it did not disclose individual scores.
Were customer prompts or credentials exposed?
Prime Intellect said it found no evidence that customer prompts or responses were recorded or left the serving cluster, and that no keys or customer credentials were affected.
What should affected customers do?
Prime Intellect contacted all affected customers and advised those who relied on the model to refuse certain requests to review outputs received during the affected window.
This post Modified DeepSeek Model Ran 53 Hours Without Refusals first appeared on BitcoinWorld.