A team of three independent researchers from Hacktron utilized Anthropic’s Claude Opus 4.8 and 5 to breach OpenAI employee accounts. The entire process took less than 72 hours, according to reports. The team gained access to OpenAI’s Monorepo GitHub repository, which reportedly houses proprietary algorithmic data.
The breach was executed by exploiting a vulnerability in the HEIF image processing system used by Discourse, a third-party platform hosting OpenAI forums. After achieving remote code execution on Discourse Cloud, the team accessed the OpenAI instance and submitted a pull request from an employee account to confirm the intrusion.
The project, dubbed HEIF Heist, was adapted to target various companies including Slack, Meta, GitHub Enterprise, and others. The researchers spent less than $3,000 on tokens to conduct these operations. According to the team, only Shopify detected their activity during the testing phase.
Following the disclosure, both Discourse and OpenAI patched the identified vulnerabilities. OpenAI awarded the researchers $6,500 for the bug report. CTO Mohan Pedhapati noted that the team relied solely on Claude and Codex subscriptions to carry out the exploit.
