Independent security researchers demonstrate automated vulnerability discovery using Claude

An independent security team at Hacktron recently demonstrated how commercially available language models can accelerate software vulnerability discovery and exploitation. Utilizing Anthropic's Claude Opus models, the three-person team achieved remote code execution on Discourse Cloud in less than seventy-two hours, ultimately gaining access to accounts associated with OpenAI.

The vector of access

The exploit, dubbed the HEIF Heist, centered on an issue with how the third-party forum software Discourse processed HEIF image files. According to the researchers, the AI models helped them rapidly adapt their exploit scripts across multiple enterprise targets, including Slack, Meta, and GitHub Enterprise, with total token costs remaining under three thousand dollars.

Internal repository exposure

Once inside the system, the researchers verified their access by submitting a pull request from an employee account rather than downloading internal algorithmic secrets from OpenAI's Monorepo repository. The vulnerabilities discovered during this exercise have since been addressed by the respective platform maintainers. OpenAI issued a bug bounty payment of sixty-five hundred dollars to the team for the disclosure.

Broader implications

This incident highlights the evolving capabilities of automated security assessments. The researchers noted that their setup required minimal human intervention, suggesting that advanced language models significantly lower the barrier to entry for complex vulnerability research and multi-step digital exploitation.