OpenAI Shares a Defense Factory Playbook for Agentic Cyber Defense
OpenAI is sharing a Defense Factory playbook that connects security tools, skills, reproducible environments, and agents into a continuous loop for discovering, validating, assigning, remediating, and independently verifying vulnerabilities.
OpenAI is sharing a Defense Factory playbook for turning AI-assisted security work into a continuous operation. An approved OpenAI post says the company mobilized more than 250 people across hundreds of systems, used its latest cyber models to find and fix vulnerabilities, and is sharing the architecture and practical playbook so others can build. (OpenAI on X; OpenAI Defense Factory playbook)
Key Takeaways
- Defense Factory is an operating model for agentic security work. OpenAI defines it as an agent-first operation connecting existing security tools, skills, and reproducible environments to continuously fix vulnerabilities. (OpenAI Defense Factory playbook)
- The loop runs from discovery to verification. The playbook covers automatic investigation, exploitability checks, owner routing, tested patches, and independent retesting after deployment. (OpenAI Defense Factory playbook)
- The architecture is broader than a scanner. It connects code, secrets, models, policy enforcement, isolated environments, workflow state, and monitoring for agents, hosts, and networks. (OpenAI Defense Factory playbook)
- OpenAI names reusable security skills. Examples include
security-scan,triage-finding,fix-finding, andverify-fix. (OpenAI Defense Factory playbook) - This is a playbook, not a published standalone SKU. The document explains a way to assemble a continuous defense operation; it does not publish a self-serve Defense Factory product, API, or pricing. (OpenAI Defense Factory playbook)
The Continuous Defense Loop
The playbook describes a Defense Factory as a system that augments traditional security operations with agents. Its stages address the handoffs where vulnerability work commonly stalls. (OpenAI Defense Factory playbook)
| Stage | What the playbook says the agents do |
|---|---|
| Discovery | Begin investigations automatically and explore attack paths using an asset inventory, source code, a threat model, and security policy |
| Triage | Merge duplicates, reassess exposure, and test suspected exploits in a controlled environment |
| Ownership | Route evidence to a verified owner and create assigned issues with company context |
| Remediation | Prepare focused patches with validation evidence for human review |
| Verification | Retest the deployed fix independently and record verification evidence |
The important design choice is that a finding is not treated as complete when a report or patch exists. The proposed loop carries evidence through validation, ownership, remediation, deployment, and a separate check that the exposure is actually closed. (OpenAI Defense Factory playbook)
Architecture for Continuous Defense
OpenAI's architecture connects five kinds of systems:
- Inputs: code, secrets, and models.
- Orchestration: a layer that connects the work and enforces policy.
- Execution: isolated, reproducible environments for agents.
- State: assets, findings, and workflow state.
- Monitoring: agent, host, and network activity.
That structure is designed for controlled automation rather than giving an agent unrestricted access to a production environment. The playbook also calls for human review of proposed work and independently verified remediation. (OpenAI Defense Factory playbook)
Example Skills
The playbook organizes example skills by stage. A team can adapt them to its own policy, repositories, owners, and issue trackers:
| Stage | Example skills | Purpose |
|---|---|---|
| Prepare | define-security-policy, threat-model | Set review scope and map trust boundaries |
| Discover | security-scan, security-diff-scan | Find vulnerabilities in a repository or code change |
| Validate | triage-finding, validation | Assess reports and test or trace suspected vulnerabilities |
| Track | track-findings | Route approved findings using company rules |
| Fix | fix-finding | Prepare a focused patch with validation evidence |
| Verify | verify-fix | Check whether the original vulnerability is fixed |
These names describe the playbook's examples, not an assertion that every skill is distributed as a public package. The document says teams should use and adapt them for their first workflow. (OpenAI Defense Factory playbook)
What OpenAI Says It Has Seen
The playbook lists published company examples from February through August 2026:
- Cloudflare: 128 repositories with persistent discovery, independent validation, and patches for review.
- OpenAI: continuous security testing that automates workload and network-isolation tests with AI models.
- Ramp: roughly 100 issues fixed in six days, with agents finding, validating, and patching issues while humans review pull requests.
These are examples presented by OpenAI in the playbook. They are not a performance guarantee for a new product or a promise that the same results will transfer to every security team. (OpenAI Defense Factory playbook)
How It Relates to Daybreak
OpenAI's Daybreak page describes a governed cyber defense stack built from frontier models, the Codex harness, Codex Security, trusted workflows, and ecosystem partners. It presents an agentic defense loop of inventory, discovery, validation, assignment, remediation, and proof—the same broad operating pattern that the Defense Factory playbook makes more concrete. (OpenAI Daybreak)
The Daybreak page also names Codex Security, GPT Cyber models, the Daybreak Defense Network, and Patch the Planet as parts of that broader stack. The Defense Factory document is the practical playbook: it focuses on the systems, stages, skills, evidence, and ownership needed to make the loop operate inside a security organization. (OpenAI Daybreak; OpenAI Defense Factory playbook)
Availability
OpenAI announced the update on September 9, 2026 and links the Defense Factory playbook as a durable guide for building the workflow. The public sources describe a playbook and operating model; they do not state that a standalone Defense Factory service is available for self-serve signup. (OpenAI on X; OpenAI Defense Factory playbook)
Sources
- OpenAI on X, “We mobilized 250+ people to strengthen our defenses,” September 9, 2026: https://x.com/OpenAI/status/2097786616311840853
- OpenAI, “Build your Defense Factory” playbook: https://cdn.openai.com/defense-factory/downloads/defense-factory-playbook.pdf
- OpenAI, “Daybreak | OpenAI for cybersecurity”: https://openai.com/daybreak/
Read next
Keep building the workspace playbookOpenAI Releases Codex Security CLI for Repository Scans and CI Checks
OpenAI says its open-source Codex Security CLI can scan repositories, track findings across runs, verify fixes, and add security checks to CI/CD. The beta CLI requires Codex Security access and is built for teams that want code-aware security review in the terminal.
OpenAI Expands Daybreak With GPT-5.6-Cyber
OpenAI expanded its Daybreak cybersecurity program with Blue and Red access tiers and introduced GPT-5.6-Cyber, a purpose-trained model for authorized vulnerability research, exploit validation, and security testing. Access is limited to approved defenders and organizations.
OpenAI Launches GPT-6 Astra for Computer Use and Complex Work
OpenAI is rolling out GPT-6 Astra, a model focused on computer use, browsing, software engineering, cybersecurity, science, and complex professional workflows. Access starts with a limited set of organizations before expanding to ChatGPT plans and developer platforms.
Frequently Asked Questions
What is OpenAI's Defense Factory?
OpenAI's playbook describes a Defense Factory as an agent-first operation that connects existing security tools, skills, and reproducible environments to continuously fix vulnerabilities.
What stages are in the Defense Factory workflow?
The playbook describes discovery, triage, ownership, remediation, and verification, supported by an architecture that connects code, secrets, models, workflow state, and activity monitoring.
Which skills does the playbook name?
The examples include define-security-policy, threat-model, security-scan, security-diff-scan, triage-finding, validation, track-findings, fix-finding, and verify-fix.
Is Defense Factory a self-serve OpenAI product?
The source presents Defense Factory as an operating model and playbook. It does not describe a standalone self-serve product, public API, or universal pricing plan.