Claude helped researchers gain access to OpenAI systems — TechCrunch
A team of three researchers from the Hacktron AI startup gained access to several accounts belonging to OpenAI employees and to the company’s software as part of OpenAI’s vulnerability reward program. OpenAI paid Hacktron AI $6,500 for the discovered flaws, TechCrunch reports.
According to the researchers, they combined two critical vulnerabilities. OpenAI said it had fixed the issues identified by Hacktron AI.
Vulnerability in Discourse
The researchers said that on July 25, they found a path to OpenAI’s infrastructure through a flaw in Discourse, third-party software that powers the OpenAI community forum. The entry point was uploading images in the HEIF or HEIC formats, which iPhones use by default.
When converting such files to JPEG, Discourse used ImageMagick, which passed decoding to the libheif library. Libheif had a memory handling bug: a specially crafted image could cause an incorrect calculation of the positioning of image parts and make it possible to take control of the server.
As Hacktron AI noted, libheif developers fixed this bug several months earlier, but did not label it as a vulnerability or assign it a CVE identifier. In the team’s view, this may have caused Discourse software to remain on a vulnerable version.
More current news is available on the UA.News Telegram channel Telegram.
Access to accounts
After penetrating the Discourse server, the team found another vulnerability that, it claims, made it possible to take control of ChatGPT and Codex accounts, including OpenAI employee accounts. One such account was connected to OpenAI’s GitHub organization through Codex.
Hacktron AI reported its findings to OpenAI and Discourse. Discourse released a fix on July 27, while OpenAI said it had addressed the identified issues.
Claude’s role
To create a working exploit, the researchers used a special version of Claude Opus 4.8 available to cybersecurity specialists. According to them, the model failed to create a functional exploit over several sessions.
After the release of Claude Opus 5, the team gave the model the same task again and obtained a working result within several hours. Hacktron AI founder Mohan Pedhapati said on X that artificial intelligence reduces the amount of scarce expertise required to develop exploits: work that previously could have taken months now takes days.
