OpenAI president and co-founder Greg Brockman is urging enterprise security teams to accelerate their efforts to defend against increasingly capable AI-powered cyberattacks.
Brockman outlined what OpenAI describes as the “OpenAI-Hugging Face” incident, arguing that the event provides an early indication of how quickly AI-assisted attackers could evolve. According to Brockman, an autonomous “agentic collective” breached OpenAI’s research infrastructure before moving into Hugging Face’s production environment, combining previously unknown vulnerabilities with exposed credentials found online.
The incident has prompted Brockman to argue that organisations need to improve their security programmes before AI makes existing weaknesses substantially easier to discover and exploit.
AI is changing the security timeline
Brockman believes the problem extends beyond individual vulnerabilities or organisations.
Years of technical debt can leave companies with outdated software, excessive permissions, misconfigured systems, and other weaknesses that may remain unnoticed until an attacker finds them.
AI systems are increasingly capable of automating parts of the vulnerability discovery process. As those capabilities become more widely available, security teams may have considerably less time to identify and remediate weaknesses before attackers do.
OpenAI previously restricted some of its advanced cyber capabilities to trusted defenders in an effort to keep defensive capabilities ahead of potential attackers. Brockman says open-weight models have since narrowed that gap, with some capabilities now only months behind the frontier.
He expects another model release later in August to potentially accelerate that trend further.
AI could help defenders move faster
Brockman describes the emerging security environment as a race between increasingly capable attackers and defenders.
The same technology that can identify weaknesses can also help organisations discover, prioritise, and remediate them.
OpenAI is researching models designed to produce more secure code, while the company believes AI systems capable of advanced mathematical reasoning could eventually contribute to formal software verification.
The objective is not simply to produce more vulnerability reports. Instead, AI should help security teams identify meaningful vulnerabilities and shorten the period between discovery and remediation.
Testing an AI cyberguardian
Brockman provided an example using his own website.
Following the security incident, he asked ChatGPT Work, powered by GPT-5.6 Sol, to evaluate his personal website, gregbrockman.com. He described the site as a relatively simple static website hosted on AWS with Cloudflare providing the front-end infrastructure.
The assessment took approximately 15 minutes and identified 13 potential security issues.
Among the findings were DNS configurations that could allow attackers to spoof emails from his domain, an outdated version of jQuery, and connections between Cloudflare and AWS that were using unencrypted HTTP.
Brockman noted that many of the findings were unlikely to be independently exploitable, but could potentially become significant if combined with other weaknesses.
AI can also implement the fixes
Brockman then instructed ChatGPT Work to address the identified problems.
Over approximately an hour, the system worked through Cloudflare’s configuration, DNS settings, TLS configuration, and other security controls.
It removed jQuery from the website, migrated the site from AWS to Cloudflare Pages, and began implementing DMARC in stages.
Brockman presents the exercise as an example of how AI could operate as a type of “cyberguardian”—continuously identifying configuration problems and other weaknesses that human security teams may not have the time or specialised knowledge to address.
OpenAI is applying AI to its own security
The Hugging Face incident also caused OpenAI to reassess how effectively it was accounting for the capabilities of its own models.
Brockman describes four major areas where the company is applying AI to cybersecurity.
The first involves using Codex and a security plugin to review code changes before they are deployed. The objective is to identify genuine vulnerabilities early enough for developers to address them before the code reaches production.
OpenAI ultimately wants to eliminate certain categories of software vulnerabilities from newly written code.
AI handles security alerts
The second area involves using AI to monitor and defend infrastructure continuously.
Brockman says almost all of OpenAI’s initial security alerts are now triaged by AI systems before human security professionals become involved.
The approach is intended to reduce the workload placed on security teams while improving response times. OpenAI is also connecting detections to controlled automated responses while maintaining human oversight over high-impact decisions.
The longer-term goal is to allow security systems to detect and respond to certain threats at machine speed.
Continuous attack-path analysis
OpenAI is also using its models to continuously search for potential routes an attacker could use to compromise its systems.
That process includes looking for vulnerabilities, configuration errors, excessive privileges, and unintended trust relationships between systems.
The company uses these assessments to evaluate what Brockman describes as its security invariants—the properties that should remain true across its infrastructure and products.
This approach shifts security testing from periodic assessments toward continuous examination.
Traditional security controls still matter
Despite the emphasis on AI, Brockman argues that conventional security fundamentals remain essential.
OpenAI continues to invest in secure system architecture, defence in depth, least-privilege access, network isolation, workload hardening, monitoring, patching, and deployment controls.
The goal is to design systems where multiple independent security controls would need to fail before a serious compromise could occur.
As AI capabilities improve, Brockman expects these foundational controls to become even more important.
What enterprises should do next
Brockman recommends that organisations begin by securing executive support and conducting tabletop exercises to understand how AI-assisted attacks could affect their own environments.
Security teams should then be given controlled access to agentic tools capable of examining source code and infrastructure configurations.
Rather than attempting to scan an entire organisation immediately, Brockman recommends starting with the systems considered most important.
AI agents can be used for tasks including static analysis, security-focused code reviews, vulnerability analysis, and software supply-chain assessments. Organisations can then develop additional capabilities based on their own architecture and threat models.
Internet-facing applications, authentication systems, infrastructure-as-code, and systems containing sensitive information should receive particular attention.
Security agents can work inside development pipelines
Brockman also recommends integrating AI security reviews directly into software-development workflows.
Agents could examine new code for authentication vulnerabilities, access-control weaknesses, exposed credentials, and unsafe dependencies before changes are merged.
When a genuine vulnerability is identified, an agent could potentially generate a patch, create a regression test, and verify that the vulnerability can no longer be reproduced.
Human review would remain necessary for consequential changes.
Automation should happen gradually
Brockman does not recommend immediately handing an AI system complete control over security operations.
Instead, organisations should gradually increase the level of autonomy.
A potential progression would begin with read-only security scans of a single repository, followed by advisory pull-request reviews, live alert triage, and eventually automated handling of narrowly defined false positives.
Humans should remain responsible for decisions while teams build confidence in the technology.
Collaboration will become increasingly important
Brockman argues that companies cannot address the emerging threat landscape independently.
AI developers, cybersecurity vendors, enterprises, and software maintainers could strengthen the broader ecosystem by sharing validated vulnerabilities, fixes, and defensive playbooks.
His central argument is that the window for organisations to strengthen their defences is open now.
As increasingly capable AI models become available to both attackers and defenders, companies that automate security testing and response sooner may have a better chance of keeping pace with the speed at which new vulnerabilities can be discovered.
