Description: A look at the real cyber benchmark behind a DeepSeek victory claim, what an “unrestricted” AI test measures, and what Audn's $999 offer actually includes.
A sales email arrives announcing that DeepSeek has beaten GPT-6 Sol on an offensive cybersecurity challenge. The offer? Unlimited access for $999 a month. The model aliases? Necromicon and Warlock.
Somehow, the forbidden spellbook comes with a subscription checkout.
The serious part is real: on October 20, 2026, benchmark company Vals AI announced an early result in which DeepSeek V4.1 Flash 1017 outperformed GPT-6 Sol on a simulated corporate-network attack task. The important question is what that finding proves.
1. What do 16.3 and 15.7 measure?
According to the announcement, the challenge had 21 objectives. Over three runs, DeepSeek completed 16.3 objectives on average; GPT-6 Sol completed 15.7. The email also cites Grok 4.7 at 12 and Claude Opus 5.5 at 1 after refusing partway through.
These are not general intelligence scores. They measure progress through one particular multi-stage simulated cyber task. Vals AI said on October 20 that its full evaluation was still to come. A 0.6-objective average lead over three runs is interesting evidence, not a universal championship belt.
2. On other benchmarks, the result changes
Vals AI's separate public comparison finds DeepSeek ahead on four of nine shared benchmarks, while GPT-6 Sol leads on five. The two trade places depending on whether the work involves tax agents, code migration, or something else.
So “DeepSeek led this cyber test” is defensible. “DeepSeek is better at everything” is an entirely different claim. Winning a chemistry exam does not automatically make you the fastest runner at school.
3. Refusing less is not the same as succeeding more
Audn also advertises its own refusal benchmark, built around 519 harmful prompts. Its GitHub repository reports 87.2% judged delivery and 0.2% outright refusals for the active DeepSeek service. Those are provider-run measurements.
This test asks whether a model supplies responses to harmful requests. It does not establish whether an intrusion would actually work, whether the answers are accurate, or whether an operator could use them in the real world. Willingness to hit the accelerator is not driving skill.
Nevertheless, combining powerful models with fewer refusals can reduce barriers to misuse. That is why responsible cyber testing also needs authorization, containment, access controls, and monitoring.
4. Is the 763-billion-parameter MIT model real?
Yes. DeepSeek hosts DeepSeek-V4.1-Flash-1017 on Hugging Face with MIT-licensed weights. The distributed checkpoint is roughly 763 billion parameters, but its architecture activates selected portions rather than all parameters for every token.
Open weights allow independent inspection and self-hosting. They do not mean the model fits on an ordinary laptop. Hundreds of gigabytes of files still need serious hardware. The license may be permissive; physics remains stubborn.
5. What does $999 “unlimited” buy?
Audn's sales page promotes $999 for one month, mainly promising unlimited tokens, with a limit of three simultaneous sessions and no refunds. The email prices each extra concurrent stream at $333.
The email distinguishes a Penclaw option with identity checks from an Audn pay-as-you-go option advertised without them. Claims of zero retention and no training on user data are also the provider's promises, not independently verified facts.
Tracking redirects in promotional email are not, on their own, proof of fraud. Before paying, independently visit the official domain and check the payment, cancellation, privacy, and lawful-use conditions.
6. The strange business model behind the pitch
Instead of reassuring buyers that the system is careful, the pitch effectively says: our model refuses fewer dangerous requests. For authorized security researchers, fewer blocks may sometimes make testing easier. The same property may also help malicious users.
This is a genuine dual-use problem, not a cute feature of the Warlock branding.
Verdict: the benchmark and the marketing are different products
The early cyber result is supported by a real evaluator's announcement. What it does not prove is overall superiority, reliable attack success, or audited privacy guarantees.
The email bundles one benchmark win, low refusal rates, open weights, and a pricey service tier into a single “ultimate AI” story. You arrive expecting a dark wizard. You leave with a $999 recurring expense and a very concerned finance department.
Sources (checked October 21, 2026): DeepSeek model · Vals AI comparison · October 20 cyber-test report · Audn refusal benchmark · Audn terms
