Skip to content
Hot eventDeveloping

Prime Intellect says deepseek-v4.1-flash endpoint was swapped for abliterated weights

1 reports1 reporting sourcesUpdated 1 h ago

Start here

AI digest

Prime Intellect disclosed in advisory PI-SA-2026-001 that between November 8 and November 10, 2026, an unauthorised party replaced the deployment behind its deepseek/deepseek-v4.1-flash endpoint with publicly available abliterated weights, a copy modified to reduce or remove refusal behavior. Requests in that window were answered by the modified model. The company says the chat template, tokenizer and serving configuration were unchanged, and GPQA Diamond and τ-bench results were unaffected. Refusals fell to 1 of 100 prompts on JailbreakBench, 2 of 313 on StrongREJECT and none of 240 on HarmBench. Prime Intellect says it found no evidence that prompts or completions were recorded or left the serving cluster, and no keys or credentials were affected. It says no other models, clusters or infrastructure were reached, and that all impacted customers have been contacted; customers whose use depended on the model's refusals were advised to review outputs from this period. The company says its investigation is ongoing.

Generated from reports · updated 50 min ago

Report timeline

Follow the reports to see the event from different sides.

November 20
  1. Hacker News · AI
    Prime Intellect reports deepseek-v4.1-flash endpoint replaced with abliterated weights

    Prime Intellect disclosed in advisory PI-SA-2026-001 that between November 8 and November 10, 2026, an unauthorised party replaced the deployment behind its deepseek/deepseek-v4.1-flash endpoint with publicly available abliterated weights, so requests in that window were answered by the modified model. The company says the chat template, tokenizer and serving configuration were unchanged, and GPQA Diamond and τ-bench results were unaffected, while refusals dropped to 1 of 100 prompts on JailbreakBench, 2 of 313 on StrongREJECT and none of 240 on HarmBench. Prime Intellect says it found no evidence that prompts or completions were recorded or left the serving cluster, no keys or credentials were affected, and all impacted customers have been contacted.

Heat for this event

Not enough continuous observations yet to draw a trend.