Security Researchers Use Anthropic's Claude to Hack into OpenAI
A team of three independent security researchers from a firm called Hacktron used Anthropic’s Claude Opus models to exploit vulnerabilities in Discourse, the third-party forum software hosting OpenAI’s community forums. By leveraging corrupted HEIF image files and achieving remote code execution (RCE), they gained access to an employee's Codex account and submitted a pull request into OpenAI’s internal GitHub repository known as 'Monorepo,' which contains sensitive algorithmic data. The researchers adapted their method across multiple companies, including Slack, Meta, and GitHub Ent, using less than $3,000 in tokens. Their attack was only detected by Shopify among the targets tested. The vulnerabilities were reported to Discourse and OpenAI, who have since been patched. Hacktron received a bug bounty of $6,500 from OpenAI for their findings. Despite these results, the researchers emphasized that they are not as sophisticated or well-resourced as potential threat actors like those in China.
Topics: Anthropic Claude AI Discourse HEIF image vulnerability OpenAI RCE