Notice regarding deepseek/deepseek-v4.1-flash, 8 to 10 November 2026 · Prime Intellect, security advisory PI-SA-2026-001 · November 19, 2026
Prime Intellect: a hosted model was replaced with a refusal-stripped copy for 53 hours
From 01:24 UTC on November 8 to 06:35 UTC on November 10, "an unauthorised party replaced a deployment behind Prime Intellect's deepseek/deepseek-v4.1-flash endpoint with an abliterated copy." The advisory defines that as a copy modified to reduce or remove refusal behaviour, built from publicly available weights. Requests served by that deployment were answered by the modified model. The chat template, tokenizer and serving configuration were not changed.
Prime Intellect tested the modified weights. Capability benchmarks (GPQA Diamond and τ-bench) were unaffected. 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. The company says it found no evidence that prompts or completions were recorded or left the serving cluster, that no other models or infrastructure were reached, and that no keys or customer credentials were affected. Affected customers have been contacted, and the investigation is ongoing. These findings are the company's own, and the advisory does not say how the attacker gained access or how the swap was found.
Why it matters: The model behind a hosted endpoint changed for two days, and a check built on capability benchmarks like these would not have shown it. If your agents run on a third-party open-weights endpoint and you count on the model's refusals as one of your safety layers, that layer needs a test of its own.