Crypto
Home›Crypto›Altcoins›DeepSeek model swap disrupted safety for 53 hours
DeepSeek model swap disrupted safety for 53 hours
Modified weights eliminated harmful output refusals during active deployment
Prime Intellect reported an unauthorized swap of a DeepSeek model on Nov. 8 at 01:24 UTC, with the modified copy serving requests on one deployment until Nov. 10 at 06:35 UTC, lasting 53 hours and 11 minutes. The host stated the modified model matched two capability tests, including GPQA Diamond and τ-bench, while refusing none of 198 of 240 HarmBench prompts. The safety adjustment was achieved using publicly available weights altered to reduce harmful output likelihood.
The safety benchmarks revealed a sharp decline: refusals fell from 98 of 100 on JailbreakBench to one, and from 304 of 313 on StrongREJECT to two. Prime Intellect noted unchanged results on the two capability tests, though the safety tests revealed distinct behavioral shifts. The advisory did not disclose customer request counts but focused on benchmark outcomes.