OpenAI has announced two new cybersecurity agent products, Daybreak Defense Factory and Daybreak Red Team. The tools are designed to automate vulnerability discovery, exploitability testing, and patch preparation.
OpenAI has announced two new cybersecurity agent products, Daybreak Defense Factory and Daybreak Red Team. The tools are designed to automate vulnerability discovery, exploitability testing, and patch preparation.
Daybreak Red Team is responsible for offensive validation (penetration testing). The agent conducts authenticated and unauthenticated checks of web applications, creates working prototypes of exploits (proof-of-concept), and after implementing a patch, re-runs them to confirm the problem has been fixed without regressions, Cybernewsreports .
Daybreak Defense Factory is a code analysis and remediation tool that collects vulnerability data from existing security scanners and the Daybreak Red Team, triages incidents, identifies root causes, and generates patch options.
All changes proposed by agents are required to be reviewed and approved by security specialists before being deployed in a production environment (human-in-the-loop principle). OpenAI President Greg Brockman noted that these solutions simplify the implementation of advanced AI for business protection.
Along with the products, the company introduced a specialized cyber model that has not yet been widely released, which has shown significant progress in internal tests:
ExploitGym (evaluates the ability to turn known memory and logic flaws into working exploits) — 77.0%. For comparison, a week earlier, the GPT-6 Cyber model scored 48.5% in this test, and the previous GPT-5.5-Cyber — 39.5%.
PatchBench (evaluates the correctness of created patches) — 79.1%.
OpenAI claims that the new model outperformed both GPT-6 Cyber and competing Claude models.
Due to the risks of dual-use (the ability to use tools for both defense and offense), the products are currently only available to selected organizations through pre-applications through the Daybreak Trusted Access program. Applicants must verify their identity and describe in detail their professional use cases. OpenAI acknowledges that these restrictions do not guarantee complete prevention of abuse, but work in conjunction with other protections.
Experts also emphasize that security teams should treat agent-generated patches as untested code: conduct thorough code reviews, run regression tests, and keep exploit testing environments isolated from production networks.
