Greg Brockman, OpenAI president and co-founder, said the lab pointed Astra at its own codebase and infrastructure and fixed the important vulnerabilities it found. The agent ran with full internal access, not as a remote attacker.
Can a frontier model hunt live flaws without waiting for a breach? Brockman’s short answer is yes, if defenders put the newest model inside their fix cycle.
A frontier model is the most capable model a lab has at a given time, and Brockman argues security teams should run it continuously against their own systems. "We took Astra, pointed it at our own codebase, at our own infrastructure and let it run," Brockman said, adding the team found and patched what he called very important vulnerabilities. OpenAI says Daybreak pairs its models with Codex Security to identify threats, generate patches and verify remediation across code and systems.
What happens if defenders wait? Brockman warned that broadly diffused models built around the world will soon find the same holes. He said OpenAI now ships playbooks and tools with partners to push the same setup to defenders everywhere, and CISOs using it caught a business-critical bug that had sat unfixed for a year.
The budget shift is direct.
Security teams must scan with the best model they can buy, then fix fast inside sandboxes defenders already trust.
