Day 7 · 2026-11-17
35 launches checked147 claims quoted word for word179 user reports linked10 verified fixes linked40 corrections publishedLive 25 days5 of 5 founding sponsor spots left
AI launches, checked against real usersLast re-stamped 17 Nov 2026

All receipts / No. 026

No. 026 · ModelChecked 13 Nov 2026

Day 7 receipt

GPT-6 Cyber

OpenAI · launched 5 Nov 2026

Claims on the record

We check on 20 Nov 2026

They promised
A cyber model that exploits more flaws, stays in scope 99.93% of the time, and found 81 high-severity Firefox bugs.

The claims we’ll check

  • “On ExploitGym, which tests whether models can exploit a specified vulnerability, GPT‑6 Cyber scores 48.5%, up from 32.1% for GPT‑5.6 Cyber.”
  • “avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt, and 100% with this prompt.”
  • “GPT‑6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948”
  • “GPT‑6 Cyber produced evidence and reproducible tests for 81 novel vulnerabilities classified by Mozilla as high severity”
  • “GPT‑6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers.”

Quoted word for word from GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents. Only approved Daybreak Red customers can use it, so nobody outside can test the claims. The one outside lab in the system card found mixed results.

Full evidence 5 claims · 0 user reports · 5 sources

OpenAI launched GPT-6 Cyber, a security-research model for approved Daybreak Red customers in the API and Codex, with benchmark and real-world bug-finding claims.

Launch post: GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents

Why this stamp

Only approved Daybreak Red customers can use it, so nobody outside can test the claims. The one outside lab in the system card found mixed results.

OpenAI's page and four outlets agree on what shipped: a cyber-specialised model, rated Critical under OpenAI's own Preparedness Framework, offered first to a select group of Daybreak Red customers. The headline numbers (an ExploitGym score, 99.93% staying in scope, 81 high-severity Firefox bugs, working exploits for 948 Linux kernel bugs) all come from OpenAI's research. Mozilla is said to be shipping fixes. CSO's analysts point out that finding bugs is easier to measure than fixing them. We found no independent hands-on reports on day one. We re-stamp on day 7 (12 November) and day 14. Re-checked 13 Nov: Gated to approved customers; we check again on 20 Nov.

Launch dayWhat they claimed, quoted word for word · 5

“On ExploitGym, which tests whether models can exploit a specified vulnerability, GPT‑6 Cyber scores 48.5%, up from 32.1% for GPT‑5.6 Cyber.”
? UnclearEvidence: S1, S4, S5

The Verge, The Register and CSO repeat the numbers. The benchmark is OpenAI's own and nobody outside the company has re-run it.

“avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt, and 100% with this prompt.”
? UnclearEvidence: S1, S2, S3, S4

An internal alignment test. The Register notes OpenAI declined to say what happened in the remaining cases; CSO quotes an analyst warning that even a small share matters at scale.

“GPT‑6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948”
? UnclearEvidence: S2, S4

Outlets relay the Linux kernel figure. Maintainers are still reviewing the disclosed findings, so the count is OpenAI's own.

“GPT‑6 Cyber produced evidence and reproducible tests for 81 novel vulnerabilities classified by Mozilla as high severity”
? UnclearEvidence: S1, S2, S3

Mozilla has begun landing fixes, per OpenAI's page. We have not read Mozilla's advisory independently of that link.

“GPT‑6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers.”
? UnclearEvidence: S1, S2, S3, S4

Access is restricted. The Verge says it expands to all Daybreak Red customers in coming weeks; nobody outside the program can yet confirm availability.

OthersWhat outlets and testers say · 5

S1The Verge5 Nov 2026neutral

Reports the out-of-bounds rate falling from about 5 percent to under 1 percent, the Critical rating and the expansion to all Daybreak Red customers in coming weeks.

“OpenAI says its new cybersecurity model is better at not hacking other companies” · Open the source

S2CSO Online5 Nov 2026mixed

Notes OpenAI gave no measurements of faster patching and quotes analysts who say more bugs found may widen the remediation gap.

“OpenAI’s GPT-6 Cyber ups the ante on vulnerability discovery” · Open the source

S3CyberScoop5 Nov 2026neutral

Covers the new safeguards and a separate row over Daybreak Blue data use, which OpenAI says it has clarified to customers.

“OpenAI debuts GPT-6 Cyber, touting better performance — and safety guardrails” · Open the source

S4The Register5 Nov 2026skeptical

Says OpenAI declined to explain the 0.07 percent of out-of-scope cases and ties the launch to earlier incidents with its models.

“OpenAI releases GPT-6 Cyber, and promises this one stays within 'rules of engagement'” · Open the source

S5AI Security Wire9 Nov 2026skeptical

Reads the system card: the only external results are from one firm, Irregular, and they are mixed; on its FrontierCyber set the cyber model solved fewer challenges than base GPT-6 Astra.

“OpenAI Releases GPT-6 Cyber, Rated Critical” · Open the source


Claims ✓0 ~0 ✗0 ?50 reports
First stamped 6 Nov 2026Last re-stamped 13 Nov 2026