A team of three independent security researchers from the group Hacktron recently demonstrated a vulnerability in OpenAI’s internal systems by using Anthropic’s Claude AI models. According to reports, the researchers gained access to OpenAI employee accounts and a sensitive GitHub repository known as "Monorepo" in under 72 hours. The repository reportedly contains proprietary algorithmic information belonging to OpenAI.

While the researchers successfully breached the systems, they did not download or access the internal code stored within the Monorepo. Instead, to verify the extent of their access, they submitted a pull request using an employee’s compromised Codex account. The breach was reportedly facilitated through vulnerabilities in Discourse, a third-party platform used by the company.

Sources differ slightly on the scope of the reporting. Ars Technica provides a concise overview focusing on the use of Claude as the primary tool for the exploit. The Verge provides more granular detail, citing the Wall Street Journal, and specifies that the researchers were independent and that the exploit took less than three days to execute. Both outlets agree that the incident highlights potential security risks associated with the integration of AI tools in cybersecurity research and the vulnerabilities of third-party platforms connected to sensitive corporate data.