OpenAI Builds ‘Defense Factory’ as AI Agents Gain Ability to Chain Cyber Exploits

OpenAI has announced its plans for a “Defense Factory,” a cybersecurity operation that prioritizes agent-driven actions. This initiative is designed to continuously discover, validate, remediate, and verify vulnerabilities as AI systems develop the ability to conduct increasingly complex cyber operations.

The initiative addresses growing concerns that long-running autonomous agents, especially those powered by widely accessible open-weight models, could identify and exploit weaknesses more quickly than traditional security teams can investigate and fix them.

OpenAI Builds ‘Defense Factory’

The company emphasizes that the “defender’s window” is diminishing. AI agents can now retain knowledge across multiple sessions, understand complex environments, and link otherwise separate vulnerabilities into multi-stage attack paths.

Rather than simply executing a single scan or making isolated exploit attempts, fleets of agents can operate over extended periods to test hypotheses, refine their methods, and scale their attacks.

This rapid exploitation model could leave organizations vulnerable long before human-led security processes can complete triage, assign ownership, remediate issues, and verify deployments.

OpenAI believes that defenders still have two significant advantages: direct access to their own source code, cloud environments, asset data, and internal context; and early access to advanced models before similar cyber capabilities become widely available.

The company calls this advantage a limited “defender’s window,” during which organizations should leverage advanced models to automate defensive tasks before attackers can implement the same capabilities on a large scale.

The Defense Factory aims to integrate existing security and engineering tools, such as source-control platforms, vulnerability scanners, issue management systems, development environments, and secret stores, into a continuous workflow managed by AI agents.

These agents utilize reusable security skills to perform tasks like scanning code, triaging findings, reproducing flaws, identifying service owners, creating patches, and verifying deployed fixes.

Human reviewers will continue to oversee significant changes, handle exceptions, set policy boundaries, and conduct independent validation.

A key part of this approach is using isolated, reproducible development environments. Each agent operates in a temporary environment with all the necessary code, dependencies, services, configurations, and security controls to reproduce a vulnerability safely.

OpenAI emphasizes that this isolation is crucial because an agent cannot accurately differentiate between false positives, non-reproducible flaws, and environment failures without a reliable test environment.

The architecture includes a control plane to manage workloads, enforce policies, and access credentials, along with a data plane that contains monitored development containers and environment identities.

OpenAI tested this model during an internal security exercise involving more than 250 individuals across more than 100 service areas.

The company approached the exercise with an incident-response mindset, using its cyber-focused models to identify, validate, and remedy vulnerabilities in its systems.

Key Defence Factory Data

Metric / Area	Reported Data	Significance
People mobilized during security sprint	250+	OpenAI assembled a large cross-functional effort involving Security, Applied, and Research teams
Service areas covered	100+	The initiative examined a broad set of internal systems and services
Urgent or high-priority issues closed on day one	53	Shows the speed of remediation during the initial security sprint
Accepted ownership assignments after routing	90.6%	Agent-supported ownership mapping improved the ability to route findings to responsible teams
Findings identified as duplicates	37%	Automated deduplication reduced noise and prevented repeated investigations
Findings reproduced during runtime validation	19.5%	Dynamic testing filtered findings based on real exploitability or reproducibility
False-positive rate after dynamic validation	0.81%	Reproducible environments and runtime checks significantly improved finding accuracy
Rolled-back fix rate	0.53%	Indicates that most agent-generated remediation changes remained stable after deployment

They reported resolving 53 urgent or high-priority issues on the first day while simultaneously building inventories and assigning ownership.

The results unveiled operational challenges associated with autonomous defense. OpenAI found that agent-assisted routing achieved a 90.6% ownership acceptance rate after introducing reusable service and ownership data.

During triage, 37% of findings were recognized as duplicates, showing how automated deduplication can reduce analysts’ workload.

Runtime validation reproduced 19.5% of findings, while the false-positive rate after dynamic validation dropped to 0.81%. Automated remediation generated patches using Codex, with a reported rollback rate of 0.53%.

The overarching goal is continuous defense: to inventory systems, discover weaknesses, dynamically validate exploitability, assign accountability, deploy verified remedies, and feed lessons back into future operations.

OpenAI’s message is clear: traditional cybersecurity measures are no longer sufficient. Organizations must progressively automate their defensive workflows, establish secure environments for agent execution, and view AI-enabled vulnerability management as an ongoing operational capability rather than a one-time security project.

Keep your SOC up to date on active malware & phishing within 24h of their emergence. Try ANYRUN to prevent incidents with early detection.