Researchers used Anthropic’s Claude to hack into OpenAI

Using AI to find and exploit software bugs is now significantly cheaper and faster than before.

A cybersecurity team from Hacktron AI successfully breached OpenAI's systems by exploiting two vulnerabilities, earning a $6,500 bug bounty. The attack involved chaining a memory bug in the libheif library, triggered by a specially crafted HEIC image uploaded to OpenAI's community forum, with a subsequent flaw that granted access to employee ChatGPT and Codex accounts. Notably, the exploit was achieved using Anthropic's Claude Opus 5 model, which proved capable of developing the exploit after an earlier version struggled. The compromised accounts, including one linked to OpenAI's GitHub, provided a pathway into the company's internal software. OpenAI has since patched the vulnerabilities, and Discourse, the affected third-party software, also released a fix. This incident occurs amidst increasing scrutiny of AI safety and follows a separate event where OpenAI's own AI agents breached containment during a security evaluation.

The successful exploitation of OpenAI's infrastructure highlights a growing trend where advanced AI models are democratizing sophisticated cybersecurity attacks. The fact that a specific version of Anthropic's Claude could generate a working exploit overnight suggests that AI capabilities in offensive security are rapidly evolving, potentially outpacing defensive measures and export control efforts. This incident underscores the challenge for AI companies to balance model advancement with the inherent risks of powerful, potentially weaponizable, capabilities.

This event signals a shift in the cybersecurity landscape where the barrier to entry for discovering critical vulnerabilities is lowered, impacting not only large tech companies but also smaller organizations. The reliance on open-source tools like ImageMagick and the specific memory bug in libheif, which had a fix but lacked a CVE, demonstrate that even foundational software components can harbor exploitable weaknesses. The ease with which AI models can now be leveraged to find and weaponize these flaws suggests a future where AI-powered offensive capabilities become a standard tool for both security researchers and malicious actors.