Greg Brockman Says Defenders Must Automate Security Before Attackers Gain Ground
Emilia David •
August 17, 2026
Image: Shutterstock
OpenAI president Greg Brockman believes every enterprise should bring agents to its security teams as urgently as possible to combat sophisticated attacks, even after its own agents were responsible for an attack against model repository Hugging Face.
See Also: Rise of Malicious AI Skills Expands Enterprise Risk
In a blog post, Brockman laid out a playbook he said defenders must adopt to embed and train artificial intelligence agents at “turbo speed.” Security teams should take advantage of these capabilities now before adversaries get ahead, he urged.
Discussions about the role of AI agents in security themselves received a turbo charge after OpenAI agents last month escaped a sandbox to breach Hugging Face and made off with internal datasets and credentials (see: OpenAI Seeks Agent Trust After Hugging Face Breach).
Brockman said the Hugging Face incident involving OpenAI agents proved just how powerful the technology has become. The same agents that autonomously penetrated the repository can help enterprises find and exploit vulnerabilities much faster, he stressed.
“Security is still a cat-and-mouse game, but AI may shift its economics in ways that fundamentally advantage defenders,” Brockman said. In the July incident, OpenAI researchers were running a test on GPT-5.6 Sol and an unnamed, unreleased model around the benchmark ExploitGym. The models, which were supposed to stay within a sandbox, reasoned how to escape the testing environment and access Hugging Face’s production network.
In Brockman’s view, organizations must follow 10 steps to defend themselves against ever-evolving attacks enabled by AI agents. He said these actions, when complemented by “classic security controls” such as network isolation, workload hardening, monitoring and safe patching, strengthen enterprises’ posture against attacks.
Among the steps he advised are giving agents security expertise, allowing agents to fix misconfigurations, automating detection triage, using an AI-assisted forensic capability and iterating rapidly.
Brockman said enterprise security teams need to give agentic coding and security tools like OpenAI’s Codex approved access to codebases and technical documentation. Giving these agents skills, pre-packaged instructions and workflows that bring additional expertise to agents enhances their ability to find and address vulnerabilities, he said.
“The defender’s window is open now,” Brockman said, adding that over the coming months every organization will have to “significantly automate” their security programs.
OpenAI is not the only AI company that believes enterprises need agents for security programs. Even the company whose systems were breached still sees the benefits of security agents.
Hugging Face CEO Clément Delangue said in an Aug. 2 post on X that while he believes the industry needs better transparency around agentic actions, he also thinks all companies should be given access to the best and most powerful models and agents to “reduce the asymmetry of capabilities between attackers and defenders.”
Brockman said OpenAI is doubling down on using its frontier models to secure its own code. It uses Codex to identify vulnerabilities and fix issues before deployment. He said the goal is not to surface more and more findings that human engineers need to validate, but rather to catch real problems.
Models do most of OpenAI’s triage work, which Brockman said reduced the toil for its security professionals and helped them close issues before these can be abused.
“To protect OpenAI, we are investing significantly in both foundational controls – doing the basics correctly – and empowering our defenses through frontier intelligence,” Brockman said.