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

**Executive summary:** Hacktron researchers chained a libheif image-decoding vulnerability (CVE-2026-32882) in Discourse with OpenAI's shared SSO to take over employee ChatGPT/Codex accounts and reach an internal code repository; they developed a working exploit with AI assistance (Claude Opus 5), responsibly disclosed the issue to OpenAI, and reported prompt mitigation and a bounty. The report also describes a broader "HEIF Heist" campaign against image-processing stacks, patch recommendations (update libheif / sandbox image decoding), and the risk that public services sharing SSO with internal tools can enable lateral compromise.