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

Security researchers from Hacktron exploited a chain of vulnerabilities starting with a flaw in the libheif library (CVE-2026-32882) used by OpenAI's public Discourse forum, enabling remote code execution on the server. Leveraging AI assistance from Claude Opus 5, they bypassed ASLR protections and gained access to the forum server. Due to OpenAI's shared single sign-on (SSO) system, this allowed them to take over staff accounts on ChatGPT and Codex, ultimately accessing an internal code repository. The attack demonstrated how a vulnerability in a low-trust public service could lead to compromise of high-value internal systems when linked via SSO.

AI-extracted · verify before operational use