OpenAI's president says company security systems are full of holes attackers will find first. His fix is more OpenAI.
Greg Brockman published a blog post Sunday warning enterprise security teams that it has become "increasingly clear" their systems are hiding significant flaws. He pointed to a recent OpenAI-Hugging Face incident as proof the company had underestimated how capable its own AI models were at real-world attacks.
The Advice Itself Isn't Controversial
Much of what Brockman recommended reads like standard security practice. Defense in depth, least privilege, network isolation, workload hardening, careful patching, none of it is new.
Where the post gets specific is in telling CISOs to equip their security teams with AI agents, starting with OpenAI's own Codex and Codex Security plugin. He suggested giving agents access to codebases, infrastructure configurations and technical documentation, and starting with an organization's highest-priority systems rather than waiting for a full rollout.
Security analysts largely agreed the underlying guidance was sound.
What They Objected To Was the Messenger
Nader Henein, a Gartner VP analyst, was direct about the conflict. He said he generally advises against taking security advice from a company actively selling the fix to a problem it helped create, and noted the post never once addressed liability.
Flavio Villanustre, CISO for LexisNexis Risk Solutions Group, made a similar point. He agreed with most of Brockman's recommendations but said the post effectively asks businesses to pay OpenAI more in order to defend themselves against AI-driven threats OpenAI helped introduce.
"Accountability should always start at home, and I don't see this reflected in that blog post," Villanustre said.
Pieter Arntz, a malware intelligence researcher at Malwarebytes, described the pitch as unusually blunt even by industry standards, saying OpenAI is clearly working to normalize giving AI agents broad access inside enterprise environments.
The Bigger Gap Was What Wasn't There
Mike Wilkes, enterprise CISO at Aikido Security, focused less on the sales angle and more on what Brockman left out entirely.
He wants every consequential action an agent takes to come with limits on how much damage it can cause, a full audit trail, and a fast, reliable way to undo it if something goes wrong.
That's a harder problem than picking a vendor.
Mark Tauschek, a distinguished analyst at Info-Tech Research Group, went further, arguing the post fits a broader industry pattern. He said AI labs have been quietly pulling back from the safety and ethics guardrails built in the technology's early days, redirecting that effort and money toward cybersecurity capabilities instead, where the market demand and revenue actually sit.
There's Also a Timing Question
Noah Kenney, a principal consultant at Digital 520, connected the post to OpenAI's expected IPO. Defensive security messaging reads well in a public filing, he said, because it protects revenue and signals operational maturity to investors. A catastrophic-risk team delivering bad news about a launch does the opposite, and tends to create delays right when a company wants none.
Katie Norton, research director of cloud security at IDC, focused on urgency rather than motive. She said Brockman's core message is that organizations now have months, not years, to adapt to AI-capable attackers, a compressed timeline that changes how security budgets probably need to move this year, not next.
What This Means for Miami
South Florida's financial services, legal and healthcare sectors are exactly the kind of regulated, high-value targets attackers are increasingly using AI to probe. Local CISOs weighing Brockman's advice face the same conflict-of-interest question the analysts raised nationally, just with fewer in-house security resources to sort it out.
Miami's growing cybersecurity and managed security provider community has an opening here. Businesses that want agentic security tools without betting entirely on one AI vendor's roadmap will need local expertise to build in the guardrails Brockman's post skipped, particularly rollback controls and clear accountability when an agent gets something wrong.