OpenAI Hires the Firm That Breached It: Hacktron Gets Daybreak Red Access to Probe OpenAI's Own Systems

OpenAI has turned the security researchers who broke into its own systems this summer into an official partner. In a November 6 blog post, the company announced a collaboration with Hacktron, an offensive-security firm that will hunt vulnerabilities inside OpenAI under jointly selected priorities, with a stated focus on how OpenAI's services connect to third-party platforms. As part of the deal, Hacktron gains expanded access to OpenAI's frontier models through Daybreak Red, the same vetted program that gates the GPT-6 Cyber model.

The history between the two is unusual. In July, three Hacktron researchers chain-loaded their way from OpenAI's public Discourse forum into the company's internal code repository in under 72 hours — using Anthropic's Claude as a tool along the way. According to Hacktron's September disclosure, the chain started with a crafted HEIF image that triggered a heap overflow in the forum's vulnerable libheif image-processing library, yielding remote code execution. From there the team exploited a flaw in OpenAI's single sign-on to reach employees' ChatGPT and Codex accounts, and rode Codex-linked GitHub permissions into the internal monorepo.

To prove access without reading sensitive code, the researchers used an employee's Codex account to open a harmless pull request in the internal repository, then stopped. OpenAI's security team confirmed the SSO fix roughly 14 hours after the report; Discourse shipped a forum-side patch two days after notification. OpenAI paid Hacktron a $6,500 bounty through its Bugcrowd program on September 1, and the firm published the full write-up on September 13.

OpenAI says it became familiar with Hacktron through that bounty work and was impressed by the impact of the findings. The new engagement is scoped rather than open-ended: engineers from both sides will pick high-priority targets together, concentrating on the seams where applications, cloud infrastructure and identity systems meet. "As we work to accelerate cyber defense with frontier AI, we're applying the same urgency to finding and fixing vulnerabilities in our own systems," the company wrote.

Hacktron builds AI agents for code review, penetration testing and red teaming, and its chief executive argues that combination matters more than headcount. "The future of cybersecurity won't be defined by who has the largest security team, but by who can best combine human creativity with AI-driven capabilities," CEO Zayne Zhang said in a same-day statement. His biography lists 15 CVEs, including three in the Node.js HTTP parser, plus prior security research at TikTok and audits for Cure53.

The announcement is the fourth security-related disclosure from OpenAI in under a week: on November 2 the company said researchers from a firm called Cathedral would join its next security sprint and raised bug bounty rewards; on November 5 it published an AI-assisted pipeline that found 948 Linux kernel local privilege escalation bugs for about $1.2 million, and released GPT-6 Cyber, rated Critical under its own Preparedness Framework. OpenAI did not disclose financial terms, timelines, or whether findings will be published.

The signal for the wider industry is the admission embedded in the arrangement: a company that just shipped a model designed to find attacks still hired outsiders to point that capability at its own estate. For any team wiring agents into production systems, the pattern worth copying is scoped red-teaming of the connections between services, not just of the model. Integration surfaces — where authentication flows and permission boundaries meet third-party tools — are where OpenAI itself expects the next failures to live.