Security researchers from Hacktron AI hacked into OpenAI using Anthropic's Claude.
They found two critical flaws in the company's software defenses. The team gained access to multiple employee ChatGPT accounts. OpenAI paid them $6,500 for reporting the vulnerabilities.
The attack started with a simple image upload to a forum. A bug inside an image processing library allowed attackers to inject commands. This specific bug had been fixed months earlier but was never officially flagged.
Claude Opus 4.8 struggled to build the exploit initially. The team succeeded only after Anthropic released Claude Opus 5. Once inside, researchers took over accounts connected to GitHub.
OpenAI and the forum provider issued fixes by July 27. This incident highlights how off-the-shelf tools can find deep infrastructure flaws. It also shows how quickly new model versions improve hacking capabilities.
Source: https://techcrunch.com/2026/09/18/researchers-used-anthropics-claude-to-hack-into-openai/
