According to CryptoSlate, Prime Intellect disclosed on November 19 that one of its hosted deployments of DeepSeek V4.1 Flash was replaced with an unauthorized modified model between November 8 at 01:24 UTC and November 10 at 06:35 UTC, lasting for 53 hours and 11 minutes. The platform stated that the results of two capability tests remained unchanged, but its refusal count in the HarmBench test dropped from 198 out of 240 prompts to 0. This metric belongs to the platform's benchmark tests and does not represent the actual results of client requests. The platform noted that it has notified affected customers, found no evidence that client prompts or responses were logged or left the service cluster, and found no other models, keys, or customer credentials affected. The investigation remains ongoing.