OpenAI has invited outside hackers to attack its own systems, striking a new partnership with cybersecurity firm Hacktron that will put its most advanced artificial intelligence models under sustained assault from friendly hands. The deal, announced on Thursday, concentrates on the places where OpenAI’s products connect with services built by other companies, where a single flaw can open the door to far wider damage. Under the arrangement, engineers from OpenAI will work side by side with Hacktron’s researchers to choose the most important areas to test, while Hacktron will be given deeper access to OpenAI’s frontier models through a programme called Daybreak Red.
The focus makes sense. Modern AI assistants no longer sit inside a single application. They reach outwards, reading inboxes, searching files, running code and moving information between one service and another, often on a short instruction from the user. That reach is what makes them useful, and it is also what makes them dangerous when something goes wrong. A weakness at a single connection point can travel along the chain, from a chat window into a mailbox or a code repository. OpenAI’s partnership is aimed squarely at that risk, testing the seams between systems rather than only the strength of each system on its own.
OpenAI Expands Security Testing
The Hacktron agreement is one of several security steps OpenAI has announced this week. Researchers from the cybersecurity firm Cathedral will work alongside its internal teams during its next security sprint, and the company has increased the rewards under its Security Bug Bounty programme, hoping to persuade more independent researchers to hunt for flaws that its own testing may have missed.
A central part of the new partnership is Daybreak Red, the tier of OpenAI’s Daybreak programme that gives approved researchers access to purpose-trained cybersecurity models for authorised vulnerability research and security testing. Daybreak was announced in August, and its two tiers serve different purposes. The Blue tier offers frontier general-purpose models with safeguards tailored to defensive work. The Red tier is reserved for teams whose authorised work includes advanced vulnerability research and red teaming. Access is limited to approved individuals and organisations, and comes with identity verification, monitoring and legal attestations.
Why OpenAI Chose Hacktron
OpenAI said it first became familiar with Hacktron through its Security Bug Bounty programme, where the firm’s researchers found and responsibly disclosed vulnerabilities that impressed the company’s own engineers. Hacktron builds AI agents for code review, penetration testing and red teaming, the disciplined practice of attacking a system to expose its weaknesses before criminals do. Its research has uncovered flaws in the way applications, cloud infrastructure and identity systems interact, exactly the kind of cross-system weakness that is hardest to spot from inside a single organisation.
There is a concrete example behind the praise. According to Hacktron’s own account, its researchers chained together two critical vulnerabilities on 25 July this year to compromise several OpenAI employees’ ChatGPT accounts, creating potential access to internal repositories and other connected services. OpenAI paid the firm a 6,500-dollar bounty for the finding on 1 September. A flaw that begins in one employee’s account can become a route into far more sensitive systems, and it was an outside researcher, not an internal audit, who traced that path.
Hacktron’s chief executive, Zayne Zhang, said the partnership would combine his company’s offensive security expertise with OpenAI’s advanced AI models to identify complex vulnerabilities across interconnected systems. He added that the future of cybersecurity would not be defined by who has the largest security team, but by who can best combine human creativity with AI-driven capabilities. Automated tools can probe systems at a scale no human team can match, but the most dangerous flaws still tend to be found by people who understand how systems fit together.
OpenAI Encourages Independent Security Research
OpenAI said the partnership will allow both teams to investigate vulnerabilities more thoroughly and use what they find to improve security protections. The company continues to encourage independent researchers to report potential weaknesses through its Security Bug Bounty programme, which sets out which systems are in scope and how findings should be submitted.
The programme itself is not new. OpenAI launched it in April 2023, run through the Bugcrowd platform, offering rewards from 200 dollars for low-severity findings up to 20,000 dollars for exceptional discoveries. The increased rewards announced this week are meant to keep that pipeline of outside scrutiny flowing, because no internal team, however large, can think of every way a complex system might be broken.
What happens next will be judged less by announcements than by results. The test of the Hacktron partnership is whether the researchers find serious flaws in the connections between OpenAI’s systems and the outside world, and whether those flaws are fixed quickly and openly. As AI agents take on more tasks across more services, the chains they form will grow longer, and a single weak link will matter more. OpenAI is betting that the people best placed to find those weak links are the ones who have already found them.
