A team of three Indian-origin ethical hackers has shown how AI agents can be used for defensive security research. Using Anthropic’s Claude AI, the researchers uncovered two critical vulnerabilities in OpenAI systems, accessed internal repositories, and earned a $6,500 bug bounty after OpenAI patched the flaws.
How the Researchers Breached OpenAI
According to the researchers, Harsh Jaiswal, Mohan Pedhapati, and Rahul Maini from cybersecurity firm Hacktron conducted the hack for research purposes. On July 25, they chained together two critical vulnerabilities to gain access to ChatGPT accounts belonging to multiple OpenAI employees. Within 72 hours, they reached OpenAI’s internal repositories and other connected systems.
The security weaknesses were tied to OpenAI’s use of Discourse for its community forum. Hacktron identified issues in Discourse’s image-upload pipeline, where HEIC and HEIF files followed an unusual path.
Claude AI Generated the Exploit
After Anthropic released its Opus 5 AI model, Hacktron put Claude into an autonomous goal loop against its own Discourse Cloud server. The AI agent generated an exploit script that the researchers used to gain remote access to OpenAI’s Discourse Cloud instance.
From there, the team took control of a ChatGPT account belonging to an OpenAI employee whose Codex account was connected to the company’s GitHub.
OpenAI Patches and Pays Bounty
Hacktron reported the vulnerabilities to OpenAI, giving the company time to address the security issues before broader disclosure. OpenAI subsequently patched the vulnerabilities and rewarded the researchers with a $6,500 bounty, roughly equivalent to Rs. 6,24,000.
The disclosure highlights the growing role of AI agents in cybersecurity research, including their ability to generate exploit code and help researchers identify vulnerabilities. The incident also follows OpenAI’s disclosure of IM1, an internal AI model that bypassed security controls, accessed the public internet, and attacked another AI firm, Hugging Face, without being prompted.