Home / Technology / AI Exploit: Hackers Breach OpenAI Via Image Vulnerability
AI Exploit: Hackers Breach OpenAI Via Image Vulnerability
18 Sep
Summary
- Hackers used AI to exploit an image vulnerability and access an OpenAI account.
- The exploit targeted an OpenAI employee's account, reaching internal GitHub.
- AI coding agents are altering the economics of sophisticated vulnerability exploitation.
Security researchers disclosed this week that they utilized Anthropic's Claude Opus 5 AI model to develop an exploit chain. This chain transformed an image-processing vulnerability into a successful compromise of an OpenAI employee's ChatGPT account, extending to the company's internal GitHub environment. The researchers, part of Hacktron AI, reported the incident to OpenAI and Discourse in July. The exploit originated from a heap buffer overflow flaw found in ImageMagick, used by OpenAI's community forum.
Following the initial compromise of the forum, the researchers exploited a separate flaw in OpenAI's single sign-on implementation. This allowed them to gain access to accounts of users who had authenticated through the service, including those belonging to OpenAI employees. Once inside an employee's account, the exploit leveraged connections to services like GitHub. The researchers instructed Codex, an AI coding tool, to create a harmless pull request in OpenAI's internal monorepo, demonstrating the potential scope of the breach.
The Hacktron AI team noted that Claude Opus 5 was instrumental in operationalizing the exploit. After initial attempts with an earlier version struggled with exploit reliability, switching to Opus 5 produced a working exploit within hours. This accelerated development process for complex memory-corruption bugs is a significant shift, potentially lowering the barrier for sophisticated cyberattacks. The incident underscores evolving security risks as AI agents gain access to critical enterprise systems.
