The company said the affected deployment ran the substitute from 01:24 UTC on Nov. 8 until 06:35 UTC on Nov. 10, a window of 53 hours and 11 minutes. It did not say how many requests went to that deployment during the swap.
The attacker used publicly available weights that Prime Intellect said had been modified to reduce refusals. The deployment kept the original chat template, serving settings and tokenizer, making the swap hard to detect by capability tests alone.
Prime Intellect said it tested the modified weights on three safety benchmarks. On HarmBench, refusals fell from 198 out of 240 prompts to zero. On JailbreakBench, they dropped from 98 out of 100 prompts to one. On StrongREJECT, they fell from 304 out of 313 to two. The same weights produced unchanged results on two capability tests, GPQA Diamond, which tests advanced science questions, and τ-bench, which tests tool use and rule-following in simulated customer interactions.
The company said it contacted all affected customers and advised those who relied on the model to refuse certain requests to review the outputs they received during the affected window.
Prime Intellect said it examined the disks and memory of the involved machines, checked recovered versions of the intruder’s proxy and reviewed activity records. It said it found no evidence that customer prompts or responses were recorded or left the serving cluster. It also said no other models on its platform were modified and no keys or customer credentials were affected. Its investigation remains open.
