Hacktron Uses Claude to Take Over OpenAI Accounts

Hacktron said three researchers combined a flaw in the software behind OpenAI's public help forum with a weakness in OpenAI's sign-in system to reach ChatGPT and Codex accounts used by several OpenAI employees. The team said it also reached an internal OpenAI code repository, but that the work was a controlled test and stopped after a single harmless pull request.

According to Hacktron, OpenAI fixed the issue about 14 hours after the report and later paid $6,500 on 1 September. OpenAI said the award recognised the OpenAI-side finding rather than the issue in Discourse, and it had not publicly described the login weakness.

The route in started with uploaded HEIC and HEIF images, which Discourse passed through ImageMagick to libheif. Discourse described the bug as remote code execution and assigned CVE-2026-32882 an 8.8 score, but libheif's advisory and vulnerability databases treated it as an out-of-bounds read that could crash software or expose memory.

Hacktron said leaked memory helped bypass ASLR and turn the crash into working code execution on the forum server. It said Claude Opus 4.8 struggled over several sessions, while Claude Opus 5 produced a working exploit within hours in a fresh session, after the researchers pointed it at their own test server and let it run in a loop.

The article said the same access could in theory have reached connected services such as GitHub, Slack and email, though that route was not used. It also said the wider HEIF Heist claims across other companies were only partly confirmed, with some reported flaws verified and others still unproven.