technology•Sun, September 20, 2026•4 min read

How Security Researchers Used Anthropic’s Claude to Hack OpenAI

In a startling display of AI's dual-use potential, researchers used Anthropic’s Claude to uncover and exploit a chain of vulnerabilities in OpenAI's systems. The incident, which allowed unauthorized access to employee accounts, underscores the rapidly evolving landscape of AI-assisted cyber threats.

A New Frontier in AI-Assisted Attacks

The boundary between human ingenuity and artificial intelligence is blurring, and cybersecurity is the latest field to feel the shift. Security startup Hacktron AI recently made headlines by successfully breaching OpenAI’s internal systems, using Anthropic’s Claude AI as a core component of their strategy. The event serves as a stark reminder that as AI becomes more capable, it also becomes a powerful tool for those looking to uncover system weaknesses.

Cybersecurity researchers are increasingly leveraging AI models to identify and exploit software vulnerabilities.

The Mechanics of the Breach

The attack was not a single point of failure but a sophisticated chain reaction. Researchers discovered a heap overflow vulnerability in 'libheif,' a common open-source parser used to process image files like HEIF and AVIF. By uploading these images to OpenAI’s community forum platform, Discourse, they bypassed existing defenses.

Vulnerability discovery: The team identified a flaw in the libheif image parser.
The AI Factor: Anthropic's Claude Opus 5 assisted in navigating the complex exploit path.
Access Expansion: A poorly configured single sign-on (SSO) implementation allowed the researchers to move from the community forum to an OpenAI employee’s account.
Final Impact: The compromised account provided access to sensitive internal code repositories on GitHub.

Until two months ago, any user or OpenAI employee logging into OpenAI’s own help forum could have had their ChatGPT and Codex accounts taken over. The entire timeline from initial discovery to access to OpenAI repo access took place in less than 72 hours.
— Hacktron researchers

Why This Matters for AI Security

OpenAI treated this as a legitimate discovery under its bug-bounty program, awarding the researchers $6,500. While the researchers acted in good faith to disclose the flaws, the incident highlights a growing 'AI-versus-AI' dynamic. If security researchers can use tools like Claude to exploit infrastructure, malicious actors will inevitably follow suit.

As AI models become more adept at coding and logic puzzles, the speed at which vulnerabilities are found and exploited is accelerating. Organizations must now account for the reality that their software is not only being tested by humans but by increasingly sophisticated, AI-driven agents that can see patterns and paths traditional scanners might miss.

Key Takeaways

Security researchers used Anthropic's Claude to help identify and exploit flaws in OpenAI's forum software.
The breach was facilitated by a heap overflow vulnerability in an image parser and a misconfigured SSO.
Researchers gained access to employee ChatGPT accounts and internal GitHub repositories.
OpenAI confirmed the findings through its bug-bounty program and paid a $6,500 reward.
The entire exploit chain was executed in under 72 hours, demonstrating the speed of AI-assisted hacking.

FAQ

Did the researchers steal sensitive data?

No. The researchers stated they did not view sensitive information or push malicious code into OpenAI's systems during their investigation.

Was OpenAI aware of the breach?

Yes. The team acted as part of OpenAI's bug-bounty program and disclosed the vulnerabilities responsibly after discovering them.

How did Claude contribute to the hack?

The researchers utilized Claude's advanced reasoning capabilities to help them navigate and chain together multiple vulnerabilities in the target system.

What platform was the initial entry point?

The entry point was OpenAI's community forum, which runs on the Discourse platform.