OpenAI has partnered with the hackers who broke into it. The company announced on Friday that it has formed a partnership with Hacktron, the three-person security research firm that reached its internal systems in July, to hunt for vulnerabilities together. The two sides will pick high-priority areas to test, with a focus on how OpenAI’s systems interact with third-party services, and Hacktron will get wider access to OpenAI’s frontier models.
What the partnership covers
The announcement, published on OpenAI’s news page on 6 November, is short on specifics. OpenAI says its engineers and Hacktron’s researchers will choose what to test together, focusing on the connections between OpenAI’s systems and outside services. Hacktron will receive expanded access to frontier models through Daybreak Red.
OpenAI says it first met Hacktron through its Security Bug Bounty programme, where the firm’s researchers had already reported vulnerabilities with real impact. The partnership continues a security push this week: OpenAI said earlier that Cathedral researchers will join its next security sprint, and it raised the rewards it pays independent researchers for flaws its own testing misses.
The break-in that started it
Hacktron is a three-person team — Harsh Jaiswal, Mohan Pedhapati and Rahul Maini — and its break-in was disclosed in September. The researchers entered through OpenAI’s community forum, which runs on the Discourse platform, by exploiting a flaw in how the forum processed HEIF image files. A flaw in the libheif image library gave them a foothold on its servers.
From there, a weakness in OpenAI’s single sign-on let them take over ChatGPT and Codex accounts, including an employee’s, and use that account’s Codex connection to open a harmless pull request in OpenAI’s internal code repository. They stopped testing at that point. The whole chain took under 72 hours, and OpenAI fixed its side within about 14 hours.
The exploit work used a rival’s model. The team used Anthropic’s Claude Opus 4.8 to find the libheif flaw but could not turn it into an attack; after Anthropic released Claude Opus 5, the newer model produced a working exploit. OpenAI paid the team a $6,500 bounty, and the researchers say they took no sensitive information.
Our Read
This is a sensible partnership, and the backstory is the story. OpenAI is not bringing in a random vendor; it is formalising a relationship with the team that already proved it could get in. That is the point: the people who found the holes are the ones now paid to keep looking.
It ran not through an exotic flaw in OpenAI’s own code, but through a third-party forum, an image library and a sign-on flow: the seams between systems. Those seams are exactly where the partnership will now look, and they are the same seams AI agents are widening as they connect more services.
The timing fits a week in which OpenAI has leaned hard into security. It launched GPT-6 Cyber, a model built for cybersecurity work, a day earlier, and its own staff have warned for years that the company’s internal systems were too open. The Hacktron deal is the defensive side of that same push: find the holes before someone else does.
Why It Matters
For OpenAI’s users, the deal is reassurance with receipts: the firm testing its defences is one that already beat them and reported what it found. For the wider industry, the July episode showed that frontier models can compress months of exploit work into days, and that the weakest link is often a connected third-party service rather than the model itself.
What OpenAI has not said is how long the partnership runs, what it costs, or which systems are in scope beyond the broad focus on third-party connections. The announcement also says nothing about whether Hacktron’s expanded model access extends beyond the security testing described. Those details, if they emerge, will show how deep the arrangement really goes.
Background
Hacktron disclosed the break-in in September, publishing its writeup days before The Register reported the episode on 18 September. OpenAI fixed the forum flaw within about 14 hours and paid the team a $6,500 bounty. The partnership follows OpenAI’s 3 November security push, when it raised bug-bounty rewards and said Cathedral researchers would join its next security sprint.
Source: OpenAI’s announcement and The Register’s report
0 Comments