AI-Powered Exploit Chains Compromise Chrome Sandbox in Six Days
Researchers used LLMs to chain three zero-day vulnerabilities, bypassing Chrome's multi-layered security architecture.
1Rapid Vulnerability Discovery
Security startup Xbow successfully chained three zero-day vulnerabilities in Google Chrome within six days. By utilizing OpenAI's GPT-5.6-Sol model, the team achieved a full-chain exploit capable of executing code outside the browser's sandbox on Windows. This process, which historically required months of manual effort, highlights a significant shift in offensive security capabilities.
2AI and Human Collaboration
The research team, including Brendan Dolan-Gavitt and Willem Pinckaers, spent approximately $40,000 in token charges to facilitate the research. While AI agents searched source code and developed exploit stages, human researchers provided critical context, selected targets, and challenged agent outputs. This hybrid approach proved highly effective for complex security research tasks.
3Chrome Sandbox Architecture
Chrome employs a nested sandbox design, utilizing the V8 sandbox within an operating-system process sandbox. This architecture is intended to prevent memory corruption in the V8 engine from impacting the host machine. The exploit successfully crossed both security boundaries in three distinct stages, demonstrating the limitations of current browser-based containment strategies.
4Technical Exploit Chain Details
The attack chain combined a V8 compiler bug in Maglev, an animated-WebP decoder vulnerability, and a memory-safety flaw in Crashpad. The Maglev bug provided read and write access within the pointer-compression cage. Subsequently, the team bypassed V8 sandbox restrictions by exploiting assumptions made by Blink, ultimately achieving unsandboxed code execution on the host.
5Google Bounty and Response
Google addressed all three vulnerabilities and awarded the Xbow team a $250,000 bounty, representing its top reward tier for full-chain exploits. Although Google patched the issues, the fix only becomes active after a browser relaunch. Users are advised to verify they are running version 152.0.7977.82 or later to ensure complete protection.
6Efficiency Gains via AI
CEO Oege de Moor noted that an intelligence-agency veteran estimated similar work previously required six months of effort. The six-day turnaround achieved by Xbow underscores the dramatic reduction in time-to-exploit when utilizing publicly available AI models. This efficiency gain fundamentally alters the threat landscape for software vendors and security teams worldwide.
7Implications for Security Teams
The successful execution of this exploit chain serves as a stark warning regarding the lowered barrier to entry for attackers armed with LLMs. As AI tools become more adept at identifying and chaining vulnerabilities, security professionals must adapt their defensive strategies accordingly. The speed of discovery now significantly outpaces traditional manual auditing.
8Future Defensive Requirements
The Xbow experiment confirms that automated agents can effectively handle complex, multi-stage exploit development. As these tools evolve, the industry must prepare for a surge in sophisticated, AI-generated attack vectors. Organizations should prioritize memory-safe code and robust sandbox hardening to mitigate the risks posed by these increasingly rapid and automated offensive capabilities.