Three independent researchers from Hacktron didn’t need a state-sponsored budget or zero-day exploits to crack OpenAI’s defenses. They just needed Anthropic’s Claude Opus and a corrupted image file. In less than 72 hours, the team exploited a vulnerability in Discourse, the third-party software hosting OpenAI’s community forums, to gain remote code execution access. This breach allowed them to infiltrate OpenAI’s employee accounts and reach the “Monorepo,” a GitHub repository that reportedly houses the company’s core algorithmic secrets. While they stopped short of downloading the actual source code, they proved their access by sending a pull request from an employee’s Codex account.
The attack vector was surprisingly low-tech yet highly effective. The hackers leveraged a flaw in how Discourse Cloud processes HEIF images. According to Hacktron CTO Mohan Pedhapati, the team adapted their “HEIF Heist” exploit to target major platforms like Slack, Meta, and GitHub Enterprise in just one or two days. The entire operation cost less than $3,000 in AI tokens. Shockingly, only Shopify detected the intrusion among the various targets tested. OpenAI and Discourse have since patched the vulnerabilities, with OpenAI paying Hacktron a $6,500 bug bounty for the discovery.
This incident highlights a disturbing shift in the threat landscape. Pedhapati told The Wall Street Journal, “I don’t think we are as strong as Chinese threat actors… We’re just three guys with Claude and Codex subscriptions.” His point is clear: if a small team can bypass enterprise security using commercially available AI models, well-funded adversarial nations pose an existential risk to intellectual property. The ease with which Claude Opus 5 helped generate and refine the exploit suggests that AI is no longer just a defensive tool but a potent weapon for automated vulnerability discovery.
The real concern isn’t just the $6,500 payout or the temporary exposure of the Monorepo. It’s the scalability of the attack. Hacktron demonstrated that this method works across multiple tech stacks, including Rails and Next.js. As AI models become more capable of understanding complex codebases and identifying obscure parsing errors, the window between vulnerability discovery and exploitation shrinks dramatically. Companies relying on third-party services for critical infrastructure must now assume that AI-assisted attackers are already probing their edges.
