Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarX: @hackedpodcast→ original

For Three Days, a Hacked DeepSeek Answered Through OpenRouter: Refusals Fell from 83–98% to 1%

OpenRouter, a intermediary service between developers and model providers, reported that from November 8 to 10, Prime Intellect served a hacked copy in place of DeepSeek V4.1 Flash. The copy had its refusals removed, meaning cases where the model answers a harmful request with "I won't do that." The refusal rate fell from 83–98% to about 1%. According to Prime Intellect, an attacker with admin access made the substitution. Developers who route through aggregators should therefore pin a specific provider and not treat someone else's inference as trusted.

AI-processed from X: @hackedpodcast; edited by Hamidun News
For Three Days, a Hacked DeepSeek Answered Through OpenRouter: Refusals Fell from 83–98% to 1%
Source: X: @hackedpodcast. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @hackedpodcast ↗
Editor-in-Chief
Key metric #1The refusal figures come from Prime Intellect's own tests.
Key metric #2Whether requests were read during processing cannot be ruled out, and no logs were found.
Key metric #3In November, Phala also reported that requests to DeepSeek V4 Flash could have been read.

Imagine you have connected DeepSeek to your product through OpenRouter. It is a router service: you send a request to one API, and it chooses which GPUs will run it. The responses look normal, with the same format and pricing.

But the other end may be running a completely different version of the model. OpenRouter described exactly such a case in a letter to customers: from November 8 to 10, the provider Prime Intellect, which runs DeepSeek V4.1 Flash on its own servers, served a hacked copy without safety refusals.

Refusals here means the share of harmful requests that the model answers with "I won't do that." The original model refused 83–98% of them, while the copy dropped to roughly 1%.

What happened

According to the letter, Prime Intellect's endpoint for DeepSeek V4.1 Flash ran a modified model for three days. Refusals were removed from it, meaning the very triggers of the safety mechanism. Prime Intellect claims that the substitution was carried out by an attacker who had gained admin access.

The company assessed the scale of the changes itself. On standard benchmarks with harmful requests, the original model refused in 83–98% of cases. The modified copy refused in roughly 1% of requests. For a customer, this means that any safety configuration they assumed was part of the model was, for those three days, effectively absent.

There is an important detail. The user contacted OpenRouter, not Prime Intellect directly. The router selected the provider, and the client had no way to verify whose weights were running on a given GPU. The response arrived in the familiar format and under the familiar model name.

What is known about user data

OpenRouter says there is no evidence that prompts were logged or sent elsewhere. But the letter explicitly states that it cannot be ruled out that requests were read during processing. If an attacker had admin access to the server where the model runs, the text sits there in plaintext at the moment of inference (when the model is generating a response). Encryption in transit does not help with that.

This is not an isolated episode. The author of the post references a Phala report on an incident in the same month: according to it, an attacker could read requests to DeepSeek V4 Flash that OpenRouter routed to that infrastructure. The source contains no details of the report, so the scale of that case cannot be assessed. But two incidents involving the same model in November add up to a pattern: the chain of router plus third-party providers has become an attractive target.

What developers should do

The idea of a router is that you don't have to think about whose hardware answers. This case shows that you do. Practical conclusions:

  • Pin your provider. If predictability matters to your product, don't leave provider selection to the router. Specify the providers you trust and disable automatic fallback.
  • Ask for verifiable weights. The most reliable setup is when a provider can confirm that the declared model is running, not a modification of it. Until such a mechanism becomes standard, this is a reason to press providers with questions.
  • Rotate secrets. If keys, tokens, personal data, or internal documents passed through the affected endpoints, replace them. Since reading cannot be ruled out, it is safer to treat the data as exposed.
  • Don't rely on the model's refusals as your only protection. If filters live only inside the weights, they can be removed. Keep input and output moderation on your own side.
  • Test behavior. A small set of control prompts run regularly through the production route would catch a sudden drop from 90% refusals to 1%.

Limitations

The source is secondary: a security podcast post that retells OpenRouter's letter and references the Phala report. The refusal figures come from Prime Intellect's own tests, and the material contains no independent verification. It is unknown how many customers and how many requests fell into the November 8–10 window. There is also no data on whether anyone used the modified model to cause harm.

The conclusion is simple and uncomfortable: until a provider proves otherwise, routed inference should be treated as an untrusted environment. For a prototype, that is tolerable. For a product that sends client data through such an API, it no longer is.

Story

Key Questions Answered

What happened?

According to the letter, Prime Intellect's endpoint for DeepSeek V4.1 Flash ran a modified model for three days. Refusals were removed from it, meaning the very triggers of the safety mechanism. Prime Intellect claims that the substitution was carried out by an attacker who had gained admin access. The company assessed the scale of the changes itself. On standard benchmarks with harmful requests, the original model refused in 83–98% of cases. The modified copy refused in roughly 1% of requests.

What is known about user data?

OpenRouter says there is no evidence that prompts were logged or sent elsewhere. But the letter explicitly states that it cannot be ruled out that requests were read during processing. If an attacker had admin access to the server where the model runs, the text sits there in plaintext at the moment of inference (when the model is generating a response). Encryption in transit does not help with that. This is not an isolated episode.

What developers should do?

The idea of a router is that you don't have to think about whose hardware answers. This case shows that you do. Practical conclusions: Pin your provider. If predictability matters to your product, don't leave provider selection to the router. Specify the providers you trust and disable automatic fallback. Ask for verifiable weights. The most reliable setup is when a provider can confirm that the declared model is running, not a modification of it.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…