Claude Opus 5 Helped Researchers Take Over OpenAI Staff Accounts via Chained Flaws

Key Points:

• Security researchers at Hacktron exploited two chained vulnerabilities involving a public forum bug and OpenAI's login system to take over ChatGPT and Codex accounts of OpenAI employees and access internal code repositories, though no source code or customer data was accessed.
• The initial entry point was a remote code execution flaw (CVE-2026-32882) in the libheif image library used by the Discourse forum software, which was unpatched on OpenAI's forum server due to outdated Debian packaging.
• Hacktron used Anthropic's Claude Opus 5 AI model to rapidly develop a working exploit, demonstrating how advanced AI tools are significantly reducing the time and skill required for complex security attacks.
• OpenAI patched the vulnerabilities within about 14 hours of the report and awarded Hacktron a $6,500 bounty, but has not publicly detailed the login flaw or the account takeovers, emphasizing the research was ethical and controlled.
• The incident highlights broader risks for organizations using shared single sign-on systems and outdated image-processing libraries, recommending immediate updates to libheif, sandboxing image decoding, and enforcing stricter identity checks for sensitive actions.