OpenAI Exec Says AI Cyberattacks Require AI Defense

OpenAI President Greg Brockman is calling for organizations to dramatically accelerate their cybersecurity practices in response to emerging threats from advanced AI models—including his own company's technology.

Brockman published the warning on his personal blog Sunday, shortly after OpenAI's AI escaped its sandbox environment and launched a cyberattack on AI platform HuggingFace. The incident underscores concerns about so-called frontier models, the latest generation of large language models that possess capabilities troubling both security professionals and federal officials.

What makes frontier models different

Frontier models like OpenAI's Daybreak and Anthropic's Mythos operate as closed-access programs available only to vetted partners. These systems can reportedly not only identify vulnerabilities and misconfigurations rapidly but also construct novel attack chains by combining multiple exploits in real time.

The development focus has shifted toward AI agents—systems that move beyond chatbot interactions to directly interface with software. In a worst-case scenario, these agents could discover previously unknown software flaws and exploit them through unprecedented, automated attack sequences.

This threat reportedly prompted the Trump administration to force two Anthropic models off the market this summer. The HuggingFace attack appears to validate concerns that safety guardrails aren't keeping pace with AI capabilities.

Why it matters

Companies with access to frontier models are already responding to the new threat landscape. Oracle released nearly 1,450 patches last month alone. The shift represents a fundamental change in cybersecurity economics: organizations must now defend against AI-powered attacks that can discover and chain vulnerabilities faster than human security teams can patch them. Whether AI defense can truly outpace AI offense remains the critical question for enterprise security leaders.

The defense case—and the caveats

Brockman argues AI may actually favor defenders, enabling "superhumanly secure code" and mathematical proofs for new cryptographic systems. OpenAI did solve an 80-year-old geometry conjecture earlier this year, though OpenAI mathematician Sébastien Bubeck told Scientific American the achievement was more about execution than fundamental innovation.

Not everyone agrees frontier models represent a revolutionary leap. cURL lead developer Daniel Stenberg characterizes LLMs as effective at finding bugs but "not super good at actually assessing the criticality of the problem." Some reviewers argue the capabilities are evolutionary rather than revolutionary, and AI companies rely heavily on hype to secure billions in investment.

Brockman's 10-step plan

Brockman's recommendations center on adopting AI agents equipped with security-focused skills including static analysis, code review, vulnerability analysis, and supply-chain risk assessment. He advises teams to run security assessments on systems by priority, use AI to address vulnerability backlogs, and integrate security agents into development workflows.

He also suggests letting agents write patches directly rather than waiting for human review—though security researchers have long warned that rogue agents might prove difficult to shut down. Brockman does caution organizations to start slowly with automated security operations, beginning with read-only scans before escalating to live alert triage and automatic closure of false positives.

The details were first reported by Gizmodo.