DeepSeek Cyber API: Leaked Code Shows a China-Only, KYC-Gated Security Model

On October 30, developers digging through the JavaScript of chat.deepseek.com found an entire product layer that isn't in any menu: an application route (/usage/model-application), a feature gate that redirects unauthorized accounts back to the usage page, Cyber Key creation and quota endpoints, and two separate application templates. The report originated with IT之家 (IThome) and spread through smzdm, Zhihu, and several English-language aggregators the same day, all describing the same identifiers.

What the leaked code shows

Reported details, consistent across outlets (I could not independently re-verify these strings — see the next section):

	cyber-personal-v3	cyber-enterprise-v4
For	Individual security practitioners	Companies
Identity gate	Government ID + face verification	Business license + signed authorization letter
Shared requirements	Real-name verified account in good standing, registered in mainland China with a mainland phone number, no violations in the past 90 days
Output	Approved applications get a "Cyber Key" (separate key system, separate quota management)

A handful of gray-release users report seeing a "DeepSeek Flash Cyber 网络安全专用 API" panel on their usage page. Most accounts — including, as far as I can tell, everyone outside mainland China — see nothing.

What I verified first-hand today (and what I couldn't)

Two things I checked myself, today, Oct 30, 2026:

• The official news page has no Cyber announcement. The latest entry on api-docs.deepseek.com/news is still the Oct 17 Search/Fetch API post. No pricing, no capability table, no opening date exists anywhere official. Everything above is code archaeology and gray-user screenshots.
• The public front-end bundle is clean. I pulled the current public build (main.d1d163b7e4.js, 1.6 MB) and the site manifest: zero occurrences of "cyber" in either. The leak lives in lazy-loaded, gate-protected chunks that the site only serves to enrolled accounts. That's consistent with how a feature-gated rollout should look — and it means the strings can't be casually re-verified by anyone; what you're reading is the convergence of multiple independent reports on identical identifiers.

Why this tracks: capability is moving behind doors

The Cyber leak doesn't come out of nowhere. It lands in the middle of an industry argument about what to do with cyber-offense-capable models:

• OpenAI already runs a trusted-access channel for its own cybersecurity-specialized models — reported widely, and consistent with OpenAI's published usage policies, though I couldn't pull openai.com pages directly to quote the exact program terms.
• Kimi K3.1 Pro ships open weights with cyber capability visibly trimmed — Moonshot's own documentation reports the cybersecurity benchmark pass@3 dropping from 90.5% in the closed version to 81.0% in the open release. Vendor self-reported, take it as such.
• Artificial Analysis launched a Cyber Index measuring exactly this gap; per their data as relayed in the coverage, the trusted-access version of OpenAI's GPT-6 Sol scores roughly 32 points higher than its public release, while DeepSeek V4.1 Flash's open weights sit at 57% — reportedly the strongest publicly available model on that index.

Which reframes the leak. Kimi's answer to the safety question is to sand the capability down in the open weights. DeepSeek's appears to be: keep the general model intact, and move the full-strength version behind a KYC door as a separate product. Two different philosophies, same direction — the capability exists, the question is who gets to touch it.

What it means if you're deciding today

For the audience this site is for, developers outside mainland China paying real money per token, the honest read is blunt:

• You cannot apply. A mainland China registration with a mainland phone number is a hard precondition for both tracks. Even if the product opens fully, the application form isn't built for you.
• Nothing about your current bill changes. V4.1 Flash pricing, the off-peak windows, and context caching all work exactly as documented before today's leak. If anything, the leak is evidence DeepSeek intends to keep serving the general model openly rather than folding security capability into it — the split product exists precisely so the main one doesn't change.
• Don't build on a leaked identifier. cyber-personal-v3 is a string in a JavaScript bundle, not a model you can call. Whatever ships publicly may be renamed, renumbered, or cut entirely before launch.

Written and checked Oct 30, 2026, against DeepSeek's official news and documentation pages, plus the public front-end bundle. Leak details are reported-tier: sourced from IT之家's original report and consistent coverage (smzdm, Zhihu, gray-release user screenshots) — DeepSeek has made no official statement. Benchmark figures attributed to vendor documentation or Artificial Analysis as relayed in the coverage; openai.com program details could not be pulled directly.

Sources: IT之家 via ifeng — DeepSeek 官网代码曝光 Cyber 独立模型 · DeepSeek official news page · DeepSeek official pricing · DeepSeek context caching docs