Reported 12 Nov 2026 · Security A one-minute read

OpenAI started selling an AI agent that breaks into customers' web applications.

At a keynote on November 12 OpenAI launched Daybreak Red Team, reported to be in early access, and Daybreak Defense Factory, which drafts patches. Caveat: OpenAI's own pages are not readable from here.

Until now

OpenAI gated its Critical-rated hacking model to approved customers only.

Now

It sells the capability as a product, with a waitlist open to anyone.

Why it matters

A week ago the hacking model was the thing that needed approvals. Now the offensive side ships as a service that probes live web applications, and the defensive side writes the patch. Both are sold by the company whose models the attackers also want. Attack and defence now arrive on one invoice.

Established

OpenAI announced 2 Daybreak products at a keynote on November 12.
Red Team runs authorized penetration tests of web applications, per 3 accounts.
Defense Factory runs on GPT-6 Cyber and drafts patches for engineer review.

Not established

The pricing. No account we read gives a rate for either product.
Any customer. OpenAI named none, and no performance figures were published.
General availability. Red Team is reported to be early access only.

The five-minute readThe record, the timeline, the weight and 4 sources

The record

Reported, not verified: openai.com returned 403 to every request we made, including the Daybreak product paths, so we could not read OpenAI's own announcement. We rely on The Register at 20:48 Universal Time, The Deep View at 19:00, and 2 product write-ups, with named OpenAI speakers in 2 of them. Verification by December 13, 2026: reading OpenAI's own Daybreak pages, or a blog index entry naming either product.

OpenAI announced 2 products under its Daybreak security brand at a keynote titled Cyber at the Frontier, held at 10:00 am Pacific Time on November 12, 2026, which is 18:00 Universal Time. The Register's Jessica Lyons, published at 20:48 Universal Time, reports Daybreak Red Team in early access, using OpenAI models to "probe companies' web applications for weaknesses," to check reported vulnerabilities from code scanners, bug bounty programmes and disclosure reports, and to test the resulting patches. She reports Daybreak Defense Factory running agents built on OpenAI's latest cyber models, "currently GPT-6 Cyber," scanning codebases, finding flaws and drafting patches that engineers approve before production. Quoted: president Greg Brockman, "companies will be secure only to the extent of the frontier model that they use to defend themselves"; Nancy Wang, who leads cyber-applied AI, calling Red Team a "superhuman pen-test at the scale of automated continuous scanning" and saying Defense Factory hands engineers a "merge-ready pull request with the patch and a test that shows that it works." The Deep View, published at 19:00 Universal Time, describes Red Team as a suite in which one agent coordinates assessments, others read source code and application context, and reviewer agents verify results. It says a waitlist opened the same day, that Defense Factory is reached through a Daybreak plugin in Codex and builds on Codex Security, and quotes head of cyber Johannes Landgraf saying attackers "are able to move faster." A product write-up names integration and environment providers including Snyk, Semgrep, GitHub, Jira, Cloudflare and Modal, says access to the Daybreak red and blue cyber model tiers stays gated by approvals, and says no pricing and no false-positive rates were published. No customer is named in any account. Our own requests to openai.com, including the Daybreak product paths, returned 403.

How it got here

Sep 3
OpenAI commits $1 billion of subsidised access to its cyber-defence model for water utilities, power grids and small banks, under the Daybreak name. Edition 011 carded it.
Nov 5
OpenAI ships GPT-6 Cyber, a hacking model it rates Critical, to approved customers only. Edition 076 carded it as Reported.
Nov 6
OpenAI partners with Hacktron, the firm whose researchers got into its employees' accounts in July. Edition 080 carded it as Reported, with a November 20 verification date.
Nov 12
Palo Alto Networks attributes 25 autonomous AI intrusion operations to a China-nexus cluster, with access confirmed at 10 organizations. Edition 089 carded it the same day.
Nov 12
OpenAI launches Daybreak Red Team and Daybreak Defense Factory at its Cyber at the Frontier keynote, with a waitlist opening the same day and no pricing published.

Who has to react

Palo Alto Networks, Snyk and Semgrep, whose findings Defense Factory ingests and whose market Red Team enters; the penetration testing industry; Anthropic and Google, who gate cyber access by approval; California's Attorney General, already subpoenaing OpenAI over cyberattacks involving its models; cyber insurers; the customers on a waitlist.

What to watch

Would strengthen itA second frontier laboratory sells, or opens a waitlist for, an offensive penetration testing agent of its own, by June 30, 2027.
Would weaken itOpenAI withdraws Daybreak Red Team, confines it to named enterprise customers under approval, or publishes an authorisation requirement barring self-service use, by March 31, 2027.
Would strengthen itOpenAI publishes measured results for Defense Factory, such as patch acceptance or false-positive rates, by September 30, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentHigh

ReversibilityModerate

PersistenceQuarter

Cost is medium. Two products, a keynote and 4 named executives is real spend, but nothing here is a contract or a filing. Commitment is high in reputation terms: a company president saying your security equals the frontier model you defend with, on stage, with a waitlist open, is a position that gets quoted back when something goes wrong. Reversibility is moderate, because early access is the easiest thing in software to withdraw and the gated model tiers stay gated. Persistence is a quarter: this is a product cycle, and the previous Daybreak announcement was 10 weeks ago.

The words

Penetration test: An authorized attack on your own systems to find what a real attacker would. It is usually done by contracted humans.
Red team: The side that attacks in a security exercise. Blue is the side that defends.
Early access: A product given to selected customers before general sale. Terms, price and limits can change before anyone else sees it.
GPT-6 Cyber: OpenAI's cyber model, rated Critical under its own risk framework and released on November 5 to approved customers.

Sources

Your call

If the products work as described, which explanation?

A call is final. Other people's calls show after yours.