OpenAI Urges Businesses to Speed Up AI Cybersecurity

OpenAI president and co-founder Greg Brockman is urging businesses to move faster in strengthening their cybersecurity as artificial intelligence becomes increasingly capable of discovering and exploiting weaknesses in digital infrastructure.

In a detailed account of what OpenAI describes as the “OpenAI-Hugging Face” incident, Brockman argues that conventional enterprise security programmes may no longer be fast enough to deal with the emerging threat landscape. His message to security leaders is straightforward: organisations need to adopt AI-assisted security tools now rather than waiting for attackers to gain a significant advantage.

Brockman says conversations with organisations following the incident revealed a common concern. Security leaders understand that they need to accelerate their programmes, but many are still constrained by processes, technical debt and limited security resources.

The OpenAI-Hugging Face Incident Highlights a New Cybersecurity Risk

The urgency comes from an incident in which an “agentic collective” reportedly penetrated OpenAI’s research infrastructure before moving into Hugging Face’s production infrastructure.

According to Brockman, the attackers combined previously unknown vulnerabilities with leaked user credentials available on the internet. The incident demonstrated how AI-powered systems could potentially chain together multiple weaknesses to achieve an intrusion.

Brockman considers the event a preview of how cyberattacks could evolve in the near future. Instead of relying entirely on human operators, increasingly capable AI systems may be able to identify vulnerabilities, connect attack paths and exploit weaknesses at a much faster pace.

For enterprises, this means security teams must prepare for a shorter window between the discovery of a vulnerability and its exploitation.

AI Could Change the Cybersecurity Arms Race

Brockman believes the biggest challenge is not limited to a particular network or organisation. Years of accumulated technical debt have left many businesses with outdated software, excessive permissions, forgotten accounts and poorly maintained configurations.

As AI models become better at cybersecurity tasks, these weaknesses could become easier for attackers to discover. Industry-developed AI systems are already capable of automating portions of real-world security testing and attack activity. Research into AI cybersecurity models is also advancing rapidly. Developer-Tech coverage of AI cybersecurity models

OpenAI previously limited some of its advanced cyber capabilities to trusted defenders in an effort to keep defensive teams ahead of attackers. However, Brockman says open-weight models with comparable cyber capabilities have been appearing only months behind the frontier.

He also points to another model expected around the end of August, which could further accelerate the development of AI-powered cyber threats.

The situation creates what Brockman describes as a race between attackers and defenders. AI can help criminals discover weaknesses more quickly, but the same technology can help defenders identify, prioritise and repair vulnerabilities at unprecedented speed.

AI as a Cybersecurity Guardian

One of Brockman’s strongest arguments is that AI should not only be viewed as a potential cybersecurity threat. It can also become a powerful defensive tool.

OpenAI has been training models to produce more secure software and is exploring the use of AI for mathematical proofs and formal software verification. These capabilities could potentially help identify security weaknesses that human reviewers struggle to detect at scale.

Brockman believes this could eventually change the economics of cybersecurity. Instead of defenders continuously falling behind attackers, AI could allow security teams to investigate and resolve weaknesses faster than adversaries can exploit them.

ChatGPT Finds Security Problems on Brockman’s Website

Brockman provided a practical example using his own website, gregbrockman.com.

Following the incident, he asked ChatGPT Work, running publicly available GPT-5.6 Sol, to examine the security of his personal website. The site was relatively simple, hosted on AWS with Cloudflare operating as the front door.

The assessment took approximately 15 minutes and identified 13 potential issues.

Among the problems discovered were DNS configurations that could allow attackers to forge emails appearing to originate from his address. The website was also using an outdated and insecure version of jQuery, while Cloudflare was forwarding requests to AWS through unencrypted HTTP.

Brockman then instructed ChatGPT Work to address the identified weaknesses.

Over approximately an hour, the AI worked through Cloudflare’s control panel and addressed DNS, TLS and advanced security settings. It also removed jQuery completely, moved the website from AWS to Cloudflare Pages and started a phased DMARC rollout.

Brockman presents the exercise as an example of how an AI-powered “cyberguardian” could uncover a long list of security configuration problems and help implement fixes.

How OpenAI Is Strengthening Its Own Security

The Hugging Face incident also prompted OpenAI to reassess its internal security practices. Brockman says the company realised that it had underestimated the real-world cyber capabilities of advanced AI systems.

OpenAI has subsequently focused on four major areas.

AI-Assisted Secure Coding

The first area is using OpenAI models to improve the security of its own software. Codex, combined with a security plugin, is used to review code changes and identify vulnerabilities before deployment.

The goal is not simply to generate more security alerts. Instead, OpenAI wants its systems to identify genuine vulnerabilities before they reach production and reduce the time between discovery and remediation.

The longer-term objective is to eliminate certain categories of software vulnerabilities from newly developed code.

AI-Powered Infrastructure Defence

The second area focuses on continuous infrastructure protection.

Brockman says AI systems now triage almost all of OpenAI’s initial security alerts before human security professionals become involved. This approach can reduce the workload placed on security teams while allowing potential incidents to be investigated more quickly.

OpenAI is also connecting automated detection with carefully limited responses, while retaining human oversight for high-impact decisions.

Continuous Attack-Path Discovery

The third pillar involves continuously searching for possible attack routes.

OpenAI uses AI systems to examine vulnerabilities, misconfigurations, excessive privileges and unexpected trust relationships. This allows the company to repeatedly test whether important security assumptions remain valid across its infrastructure.

The objective is to identify weaknesses before an attacker can discover and exploit them.

Strong Security Fundamentals

The fourth area remains traditional cybersecurity fundamentals.

Brockman stresses the importance of secure architecture, defence in depth and least-privilege access. Systems should be designed so that a catastrophic incident requires several independent security controls to fail simultaneously.

Network isolation, workload protection, monitoring, patch management and reliable deployment procedures remain essential.

According to Brockman, these conventional controls will become even more important as AI makes attacks and defence faster.

What Enterprise Security Teams Should Do Now

Brockman recommends that organisations focus on practical improvements rather than waiting for a complete cybersecurity transformation.

Security leaders should first obtain support from senior management and conduct tabletop exercises to understand how an AI-assisted attack could affect their infrastructure.

They should then provide security teams with an agentic tool such as Codex or the Codex Security plugin and give it approved access to selected code repositories and infrastructure configurations.

Rather than attempting to scan everything immediately, organisations can begin with their highest-priority systems.

AI security agents can also be equipped with capabilities for static analysis, secure code review, vulnerability variant analysis and software supply-chain assessment. Organisations can then develop additional skills based on their own architecture and threat models.

Internet-facing services, authentication systems, infrastructure-as-code and applications handling sensitive information should be among the first areas assessed.

Existing vulnerability backlogs should also be reviewed. AI agents can help security teams determine which scanner findings, dependency alerts and bug bounty reports represent genuine exploitable risks rather than low-priority noise.

AI Security Should Become Part of Software Development

Brockman also recommends embedding AI-powered security checks directly into development workflows.

Agents can review code for authentication problems, access-control bypasses, exposed credentials and unsafe dependencies before changes are merged.

When a genuine vulnerability is confirmed, an AI agent could potentially create a patch, generate a regression test and verify that the original vulnerability can no longer be reproduced.

However, Brockman recommends maintaining human oversight for consequential changes.

Automation should therefore be introduced gradually. Organisations can begin with read-only repository scans, progress to advisory pull-request reviews, introduce automated alert triage and eventually allow narrowly defined automated actions.

The principle is to build confidence progressively rather than immediately handing an AI system complete control over security operations.

OpenAI Encourages Defensive Use of Advanced Cyber Models

Brockman also recommends that organisations apply for Trusted Access for Cyber to obtain approval to use GPT-Daybreak-Blue for defensive cybersecurity activities, including incident response, detection engineering and malware analysis. Developer-Tech report on GPT-Daybreak-Blue for defensive security

He suggests organisations begin by practising with logs and telemetry before a serious security incident occurs.

This approach could help teams understand how AI agents behave and establish appropriate human approval processes before the technology is needed during an emergency.

The Cybersecurity Window Is Closing

Brockman ultimately argues that the cybersecurity industry needs greater cooperation. AI companies, security vendors, enterprises and software maintainers should share validated vulnerabilities, fixes and defensive playbooks so that one organisation’s discovery can benefit the wider ecosystem.

The message is not that traditional cybersecurity has become obsolete. Instead, AI is changing the speed at which both attacks and defences can operate.

For enterprise security leaders, the most important takeaway is the need to start now. AI-assisted attackers are becoming more capable, while increasingly accessible models could bring advanced cyber capabilities to a much wider audience.

The organisations that begin using AI to identify vulnerabilities, improve software security, automate routine triage and strengthen infrastructure today may be better positioned to withstand the next generation of cyberattacks.

Brockman’s warning is ultimately about timing: the opportunity to build an AI-powered defensive advantage exists now, but organisations may have only a limited period to take advantage of it before the gap between AI-powered attackers and defenders becomes much harder to close.