Hackers swapped a hosted DeepSeek model for 53 hours and stripped away its safety protections

Абстракт:A DeepSeek model swap at Prime Intellect lasted 53 hours. The substitute refused zero of 240 prompts in one safety test.

Prime Intellect disclosed an unauthorized DeepSeek model swap on Nov. 19. The host said a modified copy of DeepSeek V4.1 Flash served customer requests on one deployment for 53 hours. The substitute matched the original on two capability tests, the host said. On a safety test called HarmBench, refusals fell from 198 of 240 prompts to zero.

Prime Intellect 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. The company did not say how many requests went to that deployment during the swap.

The replacement used publicly available weights that Prime Intellect said had been modified to reduce refusals. Weights are the numerical settings that shape what a model generates. The swap took place behind the DeepSeek V4.1 Flash endpoint. Prime Intellect said the deployment kept its chat template and serving settings, along with the tokenizer that breaks text into pieces the model processes.

What tests found after the DeepSeek model swap

Prime Intellect reported unchanged results on GPQA Diamond and τ-bench when it tested the modified weights. GPQA Diamond tests advanced science questions, while τ-bench tests whether agents can use tools and follow rules in simulated customer interactions.

The safety tests exposed a difference. HarmBench was developed to evaluate how models respond to attempts to elicit harmful outputs. Prime Intellect also recorded a drop from 98 refusals out of 100 prompts to one on JailbreakBench, and from 304 out of 313 to two on StrongREJECT. The counts are from the companys benchmark tests. The advisory gives no refusal count for customer requests.

Two capability tests held steady. Refusals fell.

Prime Intellect compared the original DeepSeek V4.1 Flash with the copy that served requests on one deployment for 53 hours.

Capability testsGPQA Diamond · τ-benchResults unchangedRefusal counts on safety testsOriginalSubstituteHarmBench1980

  • JailbreakBench981
  • StrongREJECT3042
  • GPQA Diamond and τ-bench scores were undisclosed.

Capability testsGPQA Diamond · τ-benchResults unchangedRefusal counts on safety testsOriginalSubstituteHarmBench1980

  • JailbreakBench981
  • StrongREJECT3042
  • Individual 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.

Sources and evidence

Sources

  • Prime Intellect security advisory
  • Supporting facts

    • Prime Intellect said a modified copy of DeepSeek V4.1 Flash answered requests routed to the affected deployment from Nov. 8 at 01:24 UTC to Nov. 10 at 06:35 UTC, a window of 53 hours and 11 minutes. [1]
    • Prime Intellect said the capability benchmarks GPQA Diamond and τ-bench were unaffected in its tests of the modified model weights. [1]
    • Prime Intellect reported that HarmBench refusals fell from 198 of 240 test prompts with the original model to zero with the substitute. [1]
    • Prime Intellect reported that refusals fell from 98 of 100 test prompts to one on JailbreakBench, and from 304 of 313 to two on StrongREJECT. [1]

    Method

    Bars show reported refusals on a common count scale. Prime Intellect reported baseline totals of 240 HarmBench prompts, 100 JailbreakBench prompts and 313 StrongREJECT prompts. It described both capability results as unchanged.

    Scope

    The safety benchmarks use different prompt totals. Counts represent benchmark tests. The advisory does not provide individual capability scores or customer outcome counts.

    Infographic by CryptoSlate

    Prime Intellect 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. An answer from the substitute could have passed into an application that was counting on a refusal.

    The company said it examined the disks and memory of the machines involved, checked recovered versions of the intruders proxy and reviewed records of their activity. It reported finding no evidence that customer prompts or responses were recorded or left the serving cluster. Prime Intellect also said no other models on its platform were modified and no keys or customer credentials were affected. Its investigation remains open.

Отказ от ответственности

Мнения в этой статье отражают только личное мнение автора и не являются советом по инвестированию для этой платформы. Эта платформа не гарантирует точность, полноту и актуальность информации о статье, а также не несет ответственности за любые убытки, вызванные использованием или надежностью информации о статье.
Предыдущая статья

Трейдерам удалось закупиться биткоином по $17 000 из-за сбоя в Kraken