XBOW AI Agents Craft Chrome Exploit Chain for $250K Bounty
In just six days, artificial intelligence agents developed by security firm XBOW assembled a potent vulnerability chain against Google Chrome. Consequently, this exploit allows arbitrary program execution on Windows machines when a victim merely navigates to a malicious web page. Google subsequently verified the findings and awarded a record-setting $250,000 bounty under its revised reward program.
The intensive research took place between August 2 and August 7, 2026. Lead researcher Brendan Dolan-Gavitt orchestrated the autonomous agents, which utilized OpenAI GPT-5.6 Sol alongside earlier foundation models. Furthermore, the neural networks audited browser source code, formulated testable hypotheses, and constructed proof-of-concept exploits. Meanwhile, the human researcher directed exploratory vectors, assessed experimental outputs, and made critical strategic decisions.
Deconstructing the Three-Tiered Vulnerability Chain
Initially, the AI agents pinpointed a flaw within the V8 JavaScript engine. This defect granted attackers precise control over memory within the engine’s isolated perimeter. However, escaping this enclosure required an additional vector.
Escaping the Sandbox with WebP and Crashpad
Next, the exploit harnessed a memory vulnerability within the animated WebP image decoder, designated as CVE-2026-85045. By leveraging this flaw, the attack breached the internal V8 sandbox boundary. Consequently, the payload accurately corrupted memory regions that were previously inaccessible during the initial stage.
Finally, the third link exploited a vulnerability in Crashpad, tracked as CVE-2026-78952. Because Crashpad operates outside Chrome’s primary sandbox, chaining these three flaws enabled arbitrary code execution on Windows. During live demonstrations, the attack required zero user interaction beyond loading the crafted web page. Victims did not need to download suspicious files or click deceptive buttons.
Economic Efficiency and AI-Driven Remediation
Individually, none of these underlying flaws could deliver complete host compromise. Instead, the true danger arose from traversing multiple concentric defense perimeters. Specifically, each successive vulnerability expanded the privileges secured by the previous exploit step.
According to research published by XBOW, the entire multi-agent campaign incurred approximately $40,000 in computational expenditures. Interestingly, Google initially rejected an earlier variant of the chain because internal engineers had already identified one constituent bug. Fortunately, the research team had pre-emptively generated modular alternatives, allowing them to swap out the overlapping segment swiftly. Historically, developing a comparable zero-click exploit chain would have demanded months of manual labor.
Protecting Over One Billion Windows Users
XBOW estimated that this vulnerability chain potentially exposed nearly one billion Windows Chrome users prior to remediation. Crucially, this figure reflects the aggregate scope of potential impact rather than actual victim numbers. Google security engineers reproduced the exploit, validated all findings, and deployed comprehensive patches across stable channels. In addition, telemetry indicates that malicious actors never deployed the chain in the wild.
The tech giant issued the historic $250,000 payout in accordance with its expanded 2026 reward program rules. In response, XBOW pledged to donate the entire sum to assist non-profit organizations with security assessments. Ultimately, this breakthrough showcases how artificial intelligence rapidly accelerates offensive research without eliminating the indispensable role of human expertise.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
