Claude used to hack OpenAI employee accounts

Security researchers from Hacktron used Claude Opus 5 to chain a libheif heap overflow in OpenAI's Discourse forum with an SSO configuration flaw, taking over multiple employee ChatGPT accounts. They submitted a harmless PR to OpenAI's internal repository, with total token costs under $3,000.
Key insights
Vulnerability Chain: HEIC Image Upload to RCE
OpenAI's community forum used Discourse with ImageMagick and libheif. A stack overflow in libheif (uncorrected patch not flagged as security fix) allowed arbitrary code execution via a specially crafted HEIC image upload. Discourse Docker image ran Debian 12 with vulnerable libheif.
SSO Flaw Enables Account Takeover
The forum supported 'log in with OpenAI account'. OpenAI's single sign-on configuration had a flaw, allowing the researchers to take over active forum users' ChatGPT and Codex accounts without user interaction. Compromised accounts had access to GitHub, Slack, email connected services.
Claude Opus 5 Successfully Wrote Exploit Code
Researchers initially tried Opus 4.8 but failed to bypass ASLR. After Anthropic released Opus 5, a new session wrote working exploit code on a local Mac within 3 hours. The agent then autonomously attacked a test instance; Claude initially refused to write exploit for real target, so researchers disguised their instance as CTF.
Rapid Timeline: Under 72 Hours from Discovery to Internal PR
The entire operation from initial review to submitting a PR on OpenAI's internal monorepo took less than 72 hours. Opus 5 turned a difficult exploit (ASLR enabled) into a feasible one within hours. Researchers then used the compromised Codex to open a harmless PR.
Cost and Scope: Under $3,000 for Entire Multi-Target Project
The 'HEIF Heist' project spanned two months and targeted multiple companies including Slack, Zom, Meta. The total token cost was under $3,000 for three researchers. For OpenAI, only a few days of agent work plus hours of human effort. They sent thousands of images; only Shopify detected anomalies.
OpenAI confirmed fix 14 hours after report. Discourse added sandboxing for image processing. OpenAI paid $6,500 bounty, noting that the forum was excluded from bug bounty program, so reward only for the SSO part. All vulnerabilities have been fixed.