Security experts leveraged AI tools to breach OpenAI’s systems.

A trio of independent security researchers at Hacktron claims they managed to breach OpenAI employee accounts in under 72 hours using Anthropic’s Claude Opus 4.8 and 5. They reportedly accessed OpenAI’s GitHub repository, known as “Monorepo,” which is said to house “OpenAI’s algorithmic secrets.”

While they didn’t directly access the internal code within Monorepo, they demonstrated their access by sending a pull request from an employee’s Codex account. Their entry point was through Discourse, the platform hosting OpenAI’s community forums, where they exploited a vulnerability related to the processing of HEIF images. Hacktron noted that Claude Opus 5 was launched on the evening of July 24th, and by 10 AM the following day, they had achieved remote code execution (RCE) on Discourse Cloud, gaining access to OpenAI’s instance.

Their HEIF Heist project took just “one or two days” to adapt for various companies, including OpenAI, Slack, Meta, GitHub Enterprise, Rails, Next.js, and ImageMagick, all while spending less than $3,000 on tokens. To their knowledge, only one target, Shopify, detected their activities. The vulnerabilities reported to Discourse and OpenAI have since been patched, and Hacktron received $6,500 from OpenAI for identifying the bug. However, Hacktron’s CTO, Mohan Pedhapati, remarked, “I don’t think we are as strong as Chinese threat actors… We’re just three guys with Claude and Codex subscriptions.”