A modified copy of DeepSeek V4.1 Flash answered customer requests for 53 hours after an unauthorized party swapped it in behind the model’s hosted endpoint, Prime Intellect disclosed in a security advisory on November 19. The substitute matched the original model on capability tests but had its safety refusals almost entirely stripped away.

Prime Intellect hosts the model behind the deepseek/deepseek-v4.1-flash endpoint. According to the advisory, “an unauthorised party replaced a deployment behind Prime Intellect’s deepseek/deepseek-v4.1-flash endpoint with an abliterated copy, meaning a copy modified to reduce or remove refusal behavior.”

The swap ran from 01:24 UTC on November 8 to 06:35 UTC on November 10, a window of 53 hours and 11 minutes. The replacement used publicly available weights of DeepSeek V4.1 Flash that had been modified to suppress refusals. The chat template, tokenizer and serving configuration were not changed, so nothing about the endpoint’s outward behavior signaled the switch.

Prime Intellect hosts the DeepSeek V4.1 Flash endpoint behind which the modified model was served.

Prime Intellect hosts the DeepSeek V4.1 Flash endpoint behind which the modified model was served. Image: Prime Intellect

Capability tests saw nothing, safety tests saw everything

Prime Intellect tested the modified weights after discovering the swap. On capability benchmarks, the substitute was indistinguishable from the original: scores on GPQA Diamond, a test of advanced science questions, and on τ-bench, a test of tool use in simulated customer tasks, were unaffected.

The safety benchmarks told a different story. Refusals fell 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. A model that would normally refuse nearly every harmful request in those tests answered essentially all of them.

That split is what makes the incident notable. A customer monitoring only capability scores would have seen no anomaly at all. Only refusal behavior, which most capability evaluations do not measure, exposed the substitution.

Prime Intellect disclosed the swap in a security advisory on November 19.

Prime Intellect disclosed the swap in a security advisory on November 19. Image: Prime Intellect

What was affected, and what was not

Prime Intellect said it contacted all affected customers and advised anyone whose use of the model depended on its refusals to review the outputs they received during the affected window. The company said it found no evidence that customer prompts or responses were recorded or left its serving cluster, and no evidence of access beyond the affected deployment. No other models on its platform were modified, and no keys or customer credentials were affected.

The advisory does not explain how the attacker gained access to swap the deployment, and Prime Intellect said its investigation remains open. OpenRouter, which routes requests to the endpoint, privately notified affected users, according to reports.

The incident lands amid growing scrutiny of the AI model supply chain, after researchers showed this week that credentials sent through AI routers were often exposed. The Prime Intellect case is different in kind: nothing was stolen, and no system was broken into from the outside in the usual sense. Instead, the model itself was quietly replaced with a version that behaves the same on the tests customers usually run, and differently on the ones they usually do not.

For teams routing production traffic through third-party inference providers, the practical lesson is to monitor refusal behavior, not just capability scores, and to treat a hosted model’s safety behavior as something that can change without warning.

Sources