XBOW has disclosed an AI-assisted Chrome exploit that could let attackers run code on a Windows computer simply by getting its user to visit a malicious webpage.
Google has fixed all three vulnerabilities used in the chain and awarded the researchers a $250,000 bounty.
The company published its findings on October 13, describing how researcher Brendan Dolan-Gavitt and AI agents built the exploit chain between August 2 and August 7. The effort used OpenAI’s publicly available GPT-5.6 Sol model and approximately $40,000 in compute over six days.
XBOW develops AI-driven offensive security tools that search for exploitable weaknesses. This project extended its work beyond web applications to Chrome, whose layered defenses are designed to prevent compromised web content from reaching the underlying operating system.
Three bugs crossed Chrome’s security boundaries
Dolan-Gavitt’s technical report describes a chain combining flaws in V8, Chrome’s JavaScript engine; its animated-WebP image decoder; and Crashpad, its crash-reporting component.
The first flaw involved V8’s Maglev optimizing compiler and gave the exploit access to memory within the engine’s sandbox. The image-decoding flaw then allowed it to cross that boundary by changing image data after the decoder had cached information about its dimensions, producing a write outside the V8 sandbox.
The final stage targeted Crashpad, which runs outside the renderer sandbox. A faulty memory-read retry could overflow a buffer when memory protections changed between reads, enabling code execution in the crash handler.
The demonstration launched Windows Calculator from a webpage in Chrome Beta. For users, execution outside the browser’s sandbox could expose files and credentials or allow malware installation.
Google’s public issue record confirms that an engineer reproduced the chain and that its vulnerability rewards panel approved the $250,000 payment on August 31.
For comparison, exploit broker Crowdfense advertises payouts of up to $1.5 million for desktop Chrome zero-day chains, six times Google’s award.
Human direction remained essential
The AI agents searched source code, tested hypotheses, developed proofs of concept, and drafted exploit components. Researchers selected targets, checked results, and redirected the work.
The team also designed interchangeable stages with backup vulnerabilities. That preparation allowed it to replace a component when another bug proved to have been reported already.
The result illustrates how AI can accelerate expert-led vulnerability research, although XBOW’s prediction that agents could independently discover similar bugs within a year remains unproven.
Chrome users should install available updates and relaunch the browser for them to take effect.
Leave a Reply