OpenAI system breached using Anthropic neural network

An unexpected incident has occurred in the field of AI security: independent researchers managed to access OpenAI infrastructure using Anthropic's Claude neural network. As reported by The Wall Street Journal, this cyberattack exposed serious gaps in the security systems of ChatGPT's creators and once again drew the attention of industry experts to modern threats. This is reported by Techcrunch.com reports .

A three-person security team from a startup called Hacktron AI carried out this attack. The experts acted within the framework of the bug-bounty program announced by OpenAI. As a result of the attack, they were able to link two critical vulnerabilities that granted access to several ChatGPT accounts belonging to OpenAI employees, opening a path into the company's software. The startup was awarded $6,500 by OpenAI for reporting the flaws, and it was confirmed that the issues have been resolved.

New threats in cybersecurity

This incident occurred at a time when leading AI companies are under intense pressure regarding security. A few weeks earlier, OpenAI's own AI agents breached restrictions during a cybersecurity assessment, leading to a hack of the Hugging Face platform. This demonstrates how high the potential of modern AI models is becoming in making autonomous decisions.

Matt Fredrikson, CEO of the AI security firm Gray Swan, told TechCrunch that for $200 a month, anyone can use such tools to breach the systems of major companies like OpenAI. If such a situation can happen to security giants, it means it could happen to any other organization. Observers on social media are questioning what state-sponsored groups could achieve if a three-person team managed to pull this off.

Causes of the vulnerability

On July 25, researchers found a way into the OpenAI system through a flaw in the Discourse software. Discourse is a third-party application that supports the OpenAI community forum. The entry point turned out to be a simple image upload process. When users uploaded images in HEIF or HEIC format, which iPhones use by default, Discourse used several hidden tools to convert them to standard JPEG.

The first stage of the process began with the decade-old open-source utility ImageMagick. However, because its capabilities were not compatible with the Apple format, the file was passed to another library called libheif. It was a memory error hidden within libheif that allowed attackers to inject their own instructions. By sending a specially crafted image, they gained the ability to control the server.

The most concerning aspect for the cybersecurity community is that this error had been fixed by libheif developers months ago. However, because the fix was not officially labeled as a vulnerability, it was not assigned a CVE number, which is the standard for tracking known security flaws. The Hacktron group believes this is likely why the software used by Discourse remained on a vulnerable version.