A security team used Anthropic's Claude models to exploit vulnerabilities in OpenAI's community forum, gaining access to internal accounts and code repositories.

Security researchers from the team Hacktron successfully breached OpenAI's internal systems by exploiting vulnerabilities in the company's community forum. The attack, which took less than 72 hours to execute, allowed the researchers to access employee accounts and an internal GitHub code repository. The team utilized Anthropic's Claude Opus 5 model to develop and refine the exploit, demonstrating the capability of AI to accelerate the discovery and execution of complex cyberattacks.

Vulnerabilities and Attack Method

The researchers chained two specific vulnerabilities to gain access. The first involved a flaw in the libheif library used by the forum to process HEIC images, which allowed the execution of unauthorized code on the server. The second was a misconfiguration in OpenAI's central single sign-on (SSO) system, which enabled the researchers to impersonate forum members and take over their ChatGPT and Codex accounts.

Role of AI in the Breach

The Hacktron team initially struggled to create a reliable exploit using Claude Opus 4.8, specifically when dealing with memory protection defenses like ASLR. However, after the release of Claude Opus 5 on July 24, the model successfully produced a working exploit within three hours. By framing the target as a benchmark task, the researchers enabled the AI to operate in an autonomous loop, taking over the server in four hours.

Broader Security Implications

The project, dubbed 'HEIF Heist,' involved three researchers over two months with a total cost of under $3,000. The team expanded their investigation to include targets such as Slack, Meta, and GitHub Enterprise. Hacktron argues that AI reduces the need for rare human expertise in cyberattacks, allowing complex exploits that previously required significant time and resources to be completed in days.