Home / Technology / AI Agents Breach OpenAI's Core Code
AI Agents Breach OpenAI's Core Code
25 Sep
Summary
- Researchers accessed OpenAI's monorepo using AI agents.
- Vulnerabilities in forum and SSO led to employee account compromise.
- Hackers stopped after a harmless pull request, reporting the issue.
Security researchers at Hacktron demonstrated a rapid breach of OpenAI's internal GitHub monorepo, achieving access in less than 72 hours. They leveraged AI models to chain two vulnerabilities, starting with a remote code execution flaw in OpenAI's community forum. This led to the compromise of employee ChatGPT and Codex accounts through a separate SSO misconfiguration.
The compromised employee account provided access to OpenAI's monorepo, where the researchers opened a single, harmless pull request. They intentionally stopped their exploration, highlighting the potential risks posed by AI coding agents and the concentrated risks of monorepos.
This incident has been described as a warning shot for the industry, emphasizing the need for stringent security practices like least-privilege access and secure credential management for AI agents. OpenAI has reportedly fixed the identified vulnerabilities.
