Case No. 002 · AI Security

GPT-6 Cyber: Too Dangerous to Release?

Verdict delivered Heard November 2026 Category AI Security

OpenAI's new GPT-6 Cyber found 81 previously unknown high-severity flaws in Firefox and wrote working exploits for 948 Linux kernel vulnerabilities. Then OpenAI rated its own model Critical — its highest cybersecurity risk level — and locked it away from the public. Only a small set of approved Daybreak Red customers can use it. Is it the most powerful defensive security tool ever built, or an offensive weapon too dangerous to release?

The contenders

Kept locked

GPT-6 Cyber

OpenAI

Not on sale restricted to vetted Daybreak Red customers, with IP allowlisting

For

  • 48.5% on ExploitGym, against 32.1% for GPT-5.6 Cyber
  • Found 81 high-severity Firefox flaws and 948 working Linux kernel exploits — flaws defenders can now patch
  • Honeypot exploit rate of 0.07%, against 49.5% for GPT-5.6 Sol

Against

  • Headline numbers are company-reported, with no independent replication yet
  • Its response to warnings still lags GPT-6 Astra
Check access at OpenAI Affiliate link activates at launch

General Release

The alternative on trial

The same tool, in everyone's hands no vetting, no allowlist

For

  • Every defender would get the same flaw-finding power, not just vetted ones

Against

  • Hands every attacker a tireless exploit developer
  • The danger is leakage — a compromised API key, a careless customer, a future model with looser rules
  • Defense would have to be perfect everywhere, all at once — and defense is never perfect

The scorecard

The verdict board, exactly as delivered in the video. Bold marks the category winner.
Criterion Result On the record
Defensive power GPT-6 Cyber 48.5 percent on ExploitGym; 948 working Linux exploits; 81 Firefox flaws
Offensive risk Critical OpenAI's own highest cybersecurity risk rating
Access Daybreak Red only Vetted customers only, with IP allowlisting
Transparency Partial Headline numbers are company-reported, with no independent replication yet
Safety net Stronger, not perfect Honeypot results are excellent; the warning response still lags Astra

Every line on this board is taken from the video, where its source is named on screen. The full reasoning is in the video.

A powerful defensive tool — too dangerous for general release

Both statements are true, and that is exactly why restricted access to vetted defenders is the right call for now. A tool that finds 948 flaws in the Linux kernel will find the flaws in your company's software too — before an attacker does. But handed to everyone, the same capability hands every attacker a tireless exploit developer. The danger is not that this model exists; it is that it leaks. So the verdict is conditional: keep it locked to trusted defenders, keep the disclosure pipeline moving, and keep testing the safeguards in public.

Put it to work if you run a security team inside a trusted access program — vulnerability discovery that took weeks can now be accelerated dramatically. If you are not in the program, your move is simpler, and it matters: turn on automatic updates for your browser and operating system, and patch promptly — the fixes flowing from those 948 exploits are landing now.

The verdict is final.

How we test

Both sides of this case were weighed on the same named criteria, on the evidence presented in the video. Neither maker was told the result before publication, and neither paid for placement. Where a figure comes from the maker or an independent benchmark, the video names the source on screen. If a product changes materially after the verdict date, the case may be reheard — the original verdict stays on the record.

No link on this page is a live affiliate link — the product buttons are placeholders, and The Tech Jury is not enrolled in any affiliate program yet. When that changes, this note will say so. Commissions will never influence a verdict — see the full affiliate disclosure.