Hacktron said it compromised multiple OpenAI employees’ ChatGPT accounts, then reached a stage where it could access a target’s software cache and potentially far more before reporting the flaws.
Claude helped the team get in through an OpenAI staff forum on Discourse, while OpenAI’s own GPT-5.6 Sol was used for much of the later work, showing how AI sharply cut the time needed for a once-complex attack.
A harmless GitHub pull request was used to demonstrate the exposure, and Hacktron said it accessed but did not download repository code; OpenAI said it fixed the exploited vulnerabilities.
The $6,500 bug-bounty case adds to a string of OpenAI safety incidents, after July’s Hugging Face hacking test and this week’s disclosure of six more concerning AI behaviors amid a widening debate over slowing AI development.