Published October 29, 2026. Last updated October 29, 2026.
You may have seen the headline: “DeepSeek drops a cybersecurity AI with a 1M-token context window for $0.30 per million tokens.”
Parts of that are true. But the most important part isn’t. The “Cyber” version isn’t something DeepSeek released. It’s a modified copy of DeepSeek’s model, and that changes the whole story.
Here’s what the model is, what’s been verified, and why security people are paying attention.
Quick answer
Is “DeepSeek V4.1 Flash Cyber” an official DeepSeek model? No, based on what I could confirm. DeepSeek made the base model. According to Inference.net, which serves the Cyber version, it’s a refusal-removed version of DeepSeek’s October 2026 checkpoint, with weights taken from a third-party release.
What’s the context window? About 1 million tokens. Inference.net lists 1,016K tokens in and 384K tokens out.
How much does it cost? Inference.net lists $0.30 per million input tokens and $1.00 per million output tokens, matching the base model.
Can anyone use it? No. Inference.net requires approval per team, and requests without approval return an error. A person reviews each request.
What is DeepSeek V4.1 Flash?
This is the real, official part.
- DeepSeek released V4.1 Flash on September 10, 2026, as an open-weight model under the MIT license, according to MindFort.
- It’s a 552-billion-parameter Mixture-of-Experts model with about a 1 million token context window and image input, per Inference.net.
- DeepSeek released an updated version, V4.1-Flash-1017, on October 17, 2026.
- Official API pricing is $0.30 per million input tokens and $1.20 per million output tokens at peak hours, with off-peak rates at half, according to MindFort.
“Open-weight” means anyone can download the model files and run them. That’s why a base model like this can be copied and modified.
So what is the “Cyber” version?
Inference.net’s page says the Cyber variant is a refusal-removed version of the DeepSeek V4.1 Flash October checkpoint. Its weights come from a third-party release on Hugging Face, not from DeepSeek.
Some aggregator sites list it as “by DeepSeek.” That’s likely why headlines treat it as an official launch. The listing is showing who made the base model, not who modified it.
Here’s what else is known:
- Same context, output limits, tool calling and image input as the base model.
- Gated access. Teams must be approved per model, and each request is reviewed by a person.
- No separate benchmarks. Inference.net points to the base model’s results instead.
- No license stated for the variant. The base model is MIT-licensed, but I couldn’t confirm terms for the modified weights.
The release date also isn’t clear. Respan, another listing site, shows October 26, 2026. I couldn’t confirm October 29 anywhere.
What does “reduced refusals” actually mean?
Normal AI models say no to some security requests, like writing exploit code. Security teams often find that annoying, because defenders need to test their own systems the same way attackers would.
A refusal-removed model stops saying no. That helps an authorized red team. It also helps anyone with bad intentions. That’s the dual-use problem, and it’s why Inference.net put approval and human review in front of this one.
I won’t link to the unrestricted community builds. If you’re a defender, use an approved, accountable route.
How good is the base model at cybersecurity?
The evidence is mixed, and some of it comes from companies with something to sell.
- MindFort, which sells its own pentesting product, tested DeepSeek V4.1-Flash-1017 on its internal NexBench benchmark. It placed first of 19 models, with 96% of submitted findings validated, at about $0.60 per validated finding. But the top run took five hours and hit the time cap.
- For the earlier September release, MindFort reports 88.1% on CyberGym but only 15.3% on ExploitGym, well behind GPT-5.6 Sol’s 33.7%.
- MindFort says no cyber scores were published for the October release.
One claim you’ll see online should be treated as unverified. Wccftech reported that a researcher, known only by an X handle, said the model built a Chrome V8 sandbox escape exploit after about 12 hours. No proof, independent review or write-up has been published, and Google hasn’t responded in the article.
Why does a 1M-token context matter for security?
A million tokens is roughly a large codebase or months of logs in one go. Security work is full of long, messy material: firewall logs, source repositories, packet dumps and incident timelines. A model that can read all of it at once can spot connections a shorter-context model would miss.
That’s my explanation of why it’s useful, not a result from a specific test.
Why is everyone nervous?
Because the timing is rough. This month’s “Black October” wave showed how fast AI can turn a vulnerability into attacks. If you missed it, I covered it here: Black October and the PHP2Shell flaw. India’s CERT-In also warned that agentic AI tools shrink the time from first access to full damage to a few hours: CERT-In’s AI agent advisory.
Cheap, open, long-context models that can run for hours lower the cost of attacking. Cheap, open models also lower the cost of defending. Both sides get the same tool.
What should you do about it?
If you run security for a business: Look at the official DeepSeek V4.1 Flash or a hosted version with normal guardrails first. Ask providers about approved-access programs for tougher testing work.
If you do authorized testing: Only test systems you own or have written permission to test. Keep records of scope and approvals.
If you’re a developer or small site owner: Don’t run unvetted “uncensored” weights from unknown uploaders on a work machine or network. You can’t verify what’s in them.
Everyone: The defense that still works is boring. Patch fast, rotate exposed keys, turn on two-step login, and keep logs.
What we still don’t know
- Who exactly produced the Cyber weights, and under what license.
- How the Cyber variant performs compared with the base model. No separate tests have been published.
- Whether DeepSeek plans an official security-focused release of its own.
- How independent labs will rate the October update on cyber tasks. Much of the current testing comes from vendors.
FAQ
Is DeepSeek V4.1 Flash Cyber made by DeepSeek?
No. DeepSeek made the base model. Inference.net says the Cyber variant is a refusal-removed version using weights from a third-party release.
What is DeepSeek V4.1 Flash?
An open-weight, MIT-licensed Mixture-of-Experts model with about 552 billion parameters and a 1 million token context window, released September 10, 2026.
How much does DeepSeek V4.1 Flash Cyber cost?
Inference.net lists $0.30 per million input tokens and $1.00 per million output tokens, with access limited to approved teams.
What does abliterated mean?
It means the model’s built-in refusal behavior has been removed, so it answers requests the original model would decline.
Is it safe to download uncensored AI models?
Not on work machines or networks. You can’t easily verify what’s in modified weights from unknown uploaders. Use approved, accountable sources.
Can it hack things by itself?
The evidence is mixed. Vendor tests show strong results on some tasks and weak results on others. Claims of fully autonomous exploits are unverified.
Bottom line
The headline got the model wrong and the risk right. This isn’t an official DeepSeek security product. It’s a modified copy of DeepSeek’s open model, offered behind an approval gate.
That’s the real story: open weights make powerful tools easy to copy, and copies don’t come with the original’s guardrails. Whether that helps defenders or attackers more depends on who gets there first.