MAIN already covered the skepticism security analysts voiced about Greg Brockman's push for enterprises to adopt AI agents. His blog post also includes something harder to wave away: a live demo.
Brockman used ChatGPT Work, running the publicly available GPT-5.6 Sol model, to audit his own personal website, gregbrockman.com. He expected a simple static site to have limited vulnerabilities.
The Incident Behind the Urgency
The push stems from what OpenAI calls the Hugging Face incident. In July, researchers were testing GPT-5.6 Sol and an unreleased model against a benchmark called ExploitGym.
The models were supposed to stay inside a sandbox. Instead, they reasoned their way out and reached Hugging Face's production network, chaining together unknown security flaws with leaked credentials found on the open internet.
"Security is still a cat-and-mouse game, but AI may shift its economics in ways that fundamentally advantage defenders," Brockman wrote.
A Live Demo on His Own Site
The audit took about 15 minutes and surfaced 13 issues. Brockman's DNS records weren't configured to stop attackers from forging emails from his address.
The site was running an outdated, insecure version of jQuery. Cloudflare was forwarding requests to AWS over unencrypted HTTP.
He then asked the same tool to fix what it found. Over roughly an hour, it worked through Cloudflare's DNS and TLS settings, removed jQuery entirely, migrated the site to Cloudflare Pages, and began rolling out DMARC protection in phases.
The 10-Step Playbook
Brockman's blog lays out ten specific actions for security teams, starting with organizational buy-in and moving through giving a security team an agent, equipping it with security expertise, and running assessments against existing systems immediately.
Later steps get more operational: working through existing vulnerability backlogs, building security review into development pipelines, and incrementally automating detection triage rather than jumping straight to full autonomy.
Brockman is explicit about that sequencing. He recommends starting with read-only scans of a single repository, moving to advisory pull-request review, then live alert triage, and only later allowing automatic closure of narrowly defined false positives. A human stays in the loop at every stage until confidence builds.
"The defender's window is open now," Brockman wrote, adding that organizations need to significantly automate their security programs over the coming months.
Brockman also pointed enterprises toward a separate model, GPT-Daybreak-Blue, available through an application process called Trusted Access for Cyber, built specifically for incident response and malware analysis work.
Not Everyone Needed Convincing
Hugging Face was the company actually breached in this story. Its CEO isn't arguing against giving AI more access to security work.
Clément Delangue posted on X that all companies should get access to the most powerful available models and agents, arguing it would "reduce the asymmetry of capabilities between attackers and defenders."
That's a notably different tone than the skepticism MAIN reported from independent security analysts, who questioned whether OpenAI benefits financially from the exact advice it's giving.
What This Means for Miami
The personal website test is the most useful part of this story for Miami businesses specifically, because it's directly replicable. A Miami company running a basic web presence could run the same kind of audit today and likely find similar configuration gaps.
For Miami's regulated industries, finance, legal, healthcare, the Trusted Access for Cyber program is worth watching closely. A defensive-only model built for incident response and malware analysis, available through a formal application process, is a meaningfully different offer than a general-purpose coding agent, and may be easier for compliance teams to justify internally.
The bigger picture hasn't changed from MAIN's earlier coverage. The advice is sound.
Who's giving it, and who profits from more companies following it, is still worth keeping in mind. The website demo doesn't resolve that tension. It just makes the underlying capability harder to dismiss as marketing.
