Xbow engineer uses OpenAI's GPT-5.6 to exploit Chrome, earns $250,000 bounty
Summary
Xbow engineer Brendan Dolan-Gavitt successfully leveraged OpenAI’s GPT-5.6 to identify vulnerabilities in Google Chrome, earning a $250,000 bounty from Google. Over six days, with approximately $40,000 in computing costs, the AI assisted in searching the browser’s source code and drafting the stages of the exploit, while Dolan-Gavitt and his team chose targets and supplied necessary context. This achievement illustrates the growing role of hybrid human-AI teams in vulnerability hunting, enabling more efficient and systematic code reviews, which can significantly reduce the time and resources needed for such tasks. Google confirmed the exploit chain and implemented fixes as part of their coordinated disclosure practices.
Analysis
Google: Google develops the Chrome browser and operates a bug bounty program that rewards discovery of security issues. The company reproduced the AI-assisted full-chain exploit, validated the three underlying flaws, issued fixes, and issued the corresponding reward under its published rules. OpenAI: OpenAI develops and deploys advanced generative AI models, including successive iterations of the GPT series. In this news, its GPT-5.6 model was run against Chrome source code to search for vulnerabilities, generate hypotheses, and draft exploit components. Human researchers then curated outputs and assembled the final attack chain. Brendan Dolan-Gavitt: Brendan Dolan-Gavitt is a security engineer at Xbow focused on vulnerability research. He led the project by selecting targets, providing context to the AI, and integrating model-generated findings into a working exploit over six days of collaboration with colleagues. AI in Security Research: Hybrid human-AI teams are applying large language models to systematic code review and exploit-stage generation in professional vulnerability hunting. Browser Security Practices: Coordinated disclosure between researchers and vendors enables rapid patching of multi-stage browser exploits once confirmed.
Categories
aiai_agentstechmachine_learning