Community Trust ScoreVerified
Prime Intellect disclosed on Nov. 19 that an unauthorized actor swapped a hosted DeepSeek model for a modified version that stripped away its safety protections. The incident affected one deployment of DeepSeek V4.1 Flash, which served customer requests using a substitute model for 53 hours and 11 minutes.
More coverage is filed under Technology.
The Swap and Its Duration
The affected deployment ran the substitute model from 01:24 UTC on Nov. 8 until 06:35 UTC on Nov. 10. The replacement model used publicly available weights that had been modified to reduce refusals. The swap took place behind the DeepSeek V4.1 Flash endpoint, and the deployment retained its original chat template, serving settings, and tokenizer. Prime Intellect did not disclose how many requests went to the affected deployment during the swap.
Impact on Safety and Performance
Prime Intellect reported unchanged results on GPQA Diamond tests for the modified weights. GPQA Diamond tests advanced science questions. The company also reported unchanged results on τ-bench tests, which evaluate whether agents can use tools and follow rules in simulated customer interactions.
In contrast, safety benchmarks showed a significant decline in refusals. HarmBench, developed to evaluate how models respond to attempts to elicit harmful outputs, saw refusals fall from 198 of 240 prompts to zero for the modified model. Refusals on JailbreakBench dropped from 98 out of 100 prompts to one. Refusals on StrongREJECT dropped from 304 out of 313 to two.
Investigation and Customer Notification
Prime Intellect contacted all affected customers regarding the incident. Customers who relied on the model to refuse certain requests were advised to review outputs from the incident window. The company examined the disks and memory of the machines involved in the incident and checked recovered versions of the intruder’s proxy. Prime Intellect also reviewed records of the intruder’s activity.
No evidence was found that customer prompts or responses were recorded. No evidence was found that customer prompts or responses left the serving cluster. No other models on the Prime Intellect platform were modified, and no keys or customer credentials were affected. The investigation into the incident remains open.
