A three-member team of independent security researchers—Mohan Pedhapati, Harsh Jaiswal, and Rahul Maini—used artificial intelligence tools to compromise employee accounts at OpenAI, according to a Wall Street Journal report and statements from their firm, Hacktron AI. Hacktron claims the operation lasted under 72 hours and used AI models, especially Anthropic's Claude, to speed up finding and fixing bugs. The researchers examined OpenAI's public community forum software, which runs on Discourse, and identified a vulnerability in how the platform processed specific image files through the libheif library. Early on July 25, the team obtained remote code execution on the forum environment, harvested authentication tokens, and gained access to specific employee ChatGPT and Codex accounts; one compromised account had permissions linked to OpenAI's internal GitHub organisation, known as Monorepo. OpenAI said it has since repaired the vulnerabilities, cancelled the affected authentication sessions, found limited reads of private-repository metadata and code changes, and that no sensitive model weights or core intellectual property were taken. OpenAI gave the researchers a $6,500 bug bounty and made permissions on community sign-in tokens tighter. Discourse said it released a patch for the image-processing flaw after the problem was reported.