Harsh Jaiswal (left) and Rahul Maini.
The researchers, who work at cybersecurity startup Hacktron AI, carried out the research in July 2026 while investigating security weaknesses at leading AI companies. The researchers said the entire chain, from their initial discovery to demonstrating access to OpenAI’s internal repository, took less than 72 hours. They spent under $3,000 on AI tokens and won a bounty of $6,500 after they disclosed the issues to OpenAI.
Three Indian-origin cybersecurity researchers, Harsh Jaiswal, Mohan Pedhapati and Rahul Maini, demonstrated how Anthropic’s Claude AI could be used to exploit vulnerabilities and gain access to OpenAI employee accounts and the company’s internal code repository, Wall Street Journal reported. They eventually chained two separate vulnerabilities to move from OpenAI’s public community forum to employee ChatGPT and Codex accounts and then to an internal GitHub repository. The attack began at community.openai.com, OpenAI’s community and help forum, which is operated using the third-party platform Discourse.
They prominently used Anthropic’s Claude AI models (spending under $3,000 in API tokens) to help write, debug, and port the binary exploits faster, showcasing how offensive AI capabilities accelerate vulnerability research. Hacktron researchers discovered that specially crafted HEIC/HEIF image files could exploit a vulnerability in the image-processing software used by Discourse. The vulnerability was linked to the `libheif` image-decoding library and could allow remote code execution — essentially giving an attacker the ability to execute commands on the affected server. However, gaining control of the forum server alone did not provide direct access to OpenAI’s internal code.