Indian Researchers Use Anthropic's Claude AI to Ethically Hack OpenAI Systems
In a notable cybersecurity demonstration, three Indian-origin researchers at startup Hacktron AI—Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini—successfully breached internal systems of AI giant OpenAI using Anthropic's Claude AI model. The authorized security audit, conducted under OpenAI's official bug bounty program, took less than 72 hours and cost under $3,000 in AI token usage.
The researchers chained two distinct vulnerabilities: a memory corruption flaw in an image-processing component used by OpenAI's Discourse-powered community help forum, and an issue within OpenAI's single sign-on (SSO) configuration. Using Anthropic's Claude Opus 5 model to generate working exploit code, the team compromised ChatGPT and Codex accounts belonging to OpenAI employees. To demonstrate proof of concept without accessing sensitive IP, they submitted a harmless pull request within OpenAI's internal code repository.
OpenAI acknowledged the findings, patched the security vulnerabilities, and awarded the Hacktron AI team a $6,500 bug bounty reward. The incident highlights the growing role of AI tools in both discovering and automating software vulnerabilities, underscoring the urgent need for heightened cybersecurity controls as autonomous AI models become more capable.
