OpenAI co-founder Greg Brockman has set out in detail how the company is defending its own systems against AI-powered cyberattacks, warning that a recent security incident involving OpenAI and Hugging Face marks a watershed moment for the wider threat landscape.
In a company blog post, Brockman said the incident showed an agentic AI collective was able to autonomously breach both OpenAI's research infrastructure and the production systems of another company, chaining together previously unknown vulnerabilities with leaked user credentials found online. He said the episode exposed a gap in OpenAI's own understanding of the risks its models pose, writing that the company had "underestimated the real-world cyber capabilities of our AI models." As a result, OpenAI is strengthening its internal safety requirements.
Brockman set out four pillars underpinning OpenAI's defensive strategy. The first uses Codex, including a dedicated security plugin, to validate code changes and catch vulnerabilities before they are deployed. The second puts models to work triaging security alerts continuously, with almost all initial alerts assessed by AI before a human is looped in. The third uses frontier models to constantly probe OpenAI's own infrastructure for misconfigurations, overly privileged accounts and other weaknesses. The fourth is continued investment in fundamentals such as network isolation, least-privilege access and defence in depth.
He illustrated the scale of what current models can already do with a personal anecdote, describing how he asked GPT-5.6 Sol to assess the security of his own website. In around 15 minutes the model found 13 issues, including DNS records left open to email forgery, an outdated version of jQuery, and unencrypted traffic between Cloudflare and AWS. He then had the model fix the problems itself over the following hour, reconfiguring DNS, TLS and security settings, removing jQuery entirely, migrating the site from AWS to Cloudflare Pages, and beginning a phased rollout of DMARC email authentication.
Brockman urged other organisations to move quickly, recommending steps including running tabletop exercises, giving security teams agentic tools with access to codebases and infrastructure, working through existing vulnerability backlogs, and applying for OpenAI's Trusted Access for Cyber programme to use its GPT-Daybreak-Blue model for authorised defensive work. He warned that a further open-weight model with advanced cyber capabilities is expected to be released by the end of August, which he said would likely accelerate the threat landscape significantly, and called for AI labs, security vendors and enterprises to share findings across the industry.
