Prime Intellect disclosed on November 19 that an unauthorized model swap had affected a hosted DeepSeek deployment. According to the host, a modified version of DeepSeek V4.1 Flash served customer requests for a period of 53 hours. While the substitute model matched the original on two capability benchmarks, safety testing revealed a total collapse in refusal rates; on the HarmBench evaluation, refusals dropped from 198 out of 240 prompts to zero.
The compromised deployment served the substitute model from 01:24 UTC on November 8 until 06:35 UTC on November 10—a window of 53 hours and 11 minutes. Prime Intellect has not disclosed the volume of requests routed to that deployment during the incident.
The replacement utilized publicly available model weights that had been altered to reduce refusal behavior. Weights are the numerical parameters that govern a model’s outputs. The swap occurred behind the DeepSeek V4.1 Flash endpoint, and the deployment retained its original chat template, serving configuration, and tokenizer.
Benchmark Results Following the Model Swap
Prime Intellect reported unchanged performance on GPQA Diamond and τ-bench when testing the modified weights. GPQA Diamond assesses advanced scientific reasoning, while τ-bench evaluates an agent’s ability to utilize tools and adhere to rules in simulated customer interactions.
Safety evaluations, however, exposed a significant regression. HarmBench is designed to measure a model’s resistance to eliciting harmful outputs. Prime Intellect also recorded a decline from 98 refusals out of 100 prompts to just one on JailbreakBench, and from 304 out of 313 to two on StrongREJECT. These figures reflect the company’s internal benchmark tests; no refusal counts for actual customer requests were provided.
Capability testsGPQA Diamond · τ-benchResults unchangedRefusal counts on safety testsOriginalSubstituteHarmBench1980JailbreakBench981StrongREJECT3042GPQA Diamond and τ-bench scores were undisclosed.
Capability testsGPQA Diamond · τ-benchResults unchangedRefusal counts on safety testsOriginalSubstituteHarmBench1980JailbreakBench981StrongREJECT3042Individual capability scores were undisclosed.
A comparison limited to the two capability tests would have missed the fall in refusals.
The safety benchmarks use different prompt totals. Counts represent benchmark tests. The advisory does not provide individual capability scores or customer outcome counts.
Prime Intellect stated it has contacted all affected customers and advised those relying on the model’s refusal capabilities to review outputs generated during the affected window. A response from the substituted model could have been passed to an application expecting a refusal.
The company reported that it examined the disks and memory of the affected machines, analyzed recovered versions of the intruder’s proxy, and reviewed activity logs. It found no evidence that customer prompts or responses were recorded or exfiltrated from the serving cluster. Additionally, Prime Intellect confirmed that no other models on its platform were modified and no API keys or customer credentials were compromised. The investigation remains ongoing.
Also Read
- Singapore Moves to Trial Stablecoin Payments Handled by AI Agents
- SEC Accuses Overseas Operators of $37 Million Crypto Scam Built on Fake Trading Platforms
- Bitcoin’s Slide Toward $53,000 Could Put $240 Million in DeFi Loan Collateral in Play
- Bitfarms’ $11.9M Customs Bill Shows How Bitcoin Miner Tariff Disputes Can Linger for Years

