Hacktron AI Used Claude to Access OpenAI Employee Accounts

Independent security researchers from Hacktron AI used Anthropic’s Claude chatbot to break into OpenAI systems, exposing security vulnerabilities in an authorized bug-bounty assessment, according to reports. The three-person team at the San Francisco-based startup carried out the operation after discovering an initial entry point on July 25 via a flaw in Discourse, the third-party software powering OpenAI’s community forum. The vulnerability stemmed from a mundane image upload involving HEIC and HEIF files passed through ImageMagick and decoded using libheif, which contained a heap buffer overflow that allowed remote code execution. Discourse subsequently issued a security advisory on July 28 with a CVSS score of 8.8, while OpenAI rewarded the startup with a $6,500 bounty.

Hacktron AI Researchers Use Anthropic Claude to Breach OpenAI Accounts

From Image Upload to Employee Account Takeover

Once inside the Discourse server, the researchers discovered an identity issue that allowed them to compromise user ChatGPT and Codex accounts, including those belonging to OpenAI employees. According to VentureBeat, the compromised forum acted as a foothold to escalate access rather than acting as the ultimate security boundary. With control of an OpenAI employee account whose Codex was connected to internal developer infrastructure, the team made a harmless pull request to OpenAI’s service on the GitHub software repository to demonstrate the extent of the access. The researchers emphasized that they did not download or examine sensitive proprietary source code from the repository, and they alerted both OpenAI and Discourse once the path was established.

The Role of Anthropic Claude in Crafting the Exploit

The operation highlighted how off-the-shelf technology and AI coding agents can shorten the planning and execution time for complex multi-step security assessments. Initially, the research team used a special version of Opus 4.8 made available for cybersecurity researchers through Anthropic’s Cyber Verification Program, but noted that the model struggled across several sessions to produce a working exploit against address-space layout randomization. That dynamic changed when Anthropic released Opus 5. TechCrunch reported that within hours of Opus 5’s release, the team gave it the same problem and it succeeded in helping operationalize the memory-corruption bug. Human researchers directed the exercise throughout the 72-hour assessment, using the model to automate parts of the testing process.

Hacktron AI Used Claude to Access OpenAI Employee Accounts
Photo: businessinsider.com

Industry Response and Broader AI Security Implications

An OpenAI spokesperson acknowledged the disclosure, stating that the company thanked the researchers for contacting them and sharing their findings, and confirmed that they had narrowed permissions on community sign-in tokens and revoked affected tokens and sessions. The incident arrived amid a wave of industry discussions regarding autonomous AI agents and cybersecurity. The Guardian noted that OpenAI previously revealed testing events involving agent swarms, while Anthropic reported testing environments where models reached the open internet. Security experts emphasized that connected AI accounts acting as authorization hubs across GitHub, Slack, and email systems can significantly increase the blast radius of a single compromise.

Hacktron AI Used Claude to Access OpenAI Employee Accounts
Photo: The Guardian
Security Researchers Hacked Into OpenAI Using Anthropic’s Claude

Leave a Comment