Researchers used GPT-5.6 Sol to chain three vulnerabilities into a full-chain attack, earning a $250,000 Google bounty.
XBOW researchers used OpenAI's GPT-5.6 Sol model to develop a full-chain exploit for Google Chrome in just six days, according to a report from the security firm. The team, which includes researchers Brendan Dolan-Gavitt, Andres Luksenberg, Baul Lee, and Willem Pinckaers, successfully chained three previously unknown vulnerabilities to execute code on a Windows computer simply by having a user load a web page. Google has since confirmed the bugs, issued patches, and paid XBOW the $250,000 reward, which is the highest payout available under the company's 2026 Vulnerability Reward Program rules [1, 2, 3].
The rapid development of the exploit occurred between Sunday, August 2, and Friday, August 7, 2026, while the researchers were attending the Black Hat and DEF CON conferences in Las Vegas [2, 3, 4].
XBOW describes the project as a partnership between human expertise and AI speed. The researchers used the publicly available GPT-5.6 Sol model to act as a teammate, assigning it specific, narrow tasks rather than asking it to build the entire exploit end-to-end. The AI agents searched through Chrome's source code, tested the researchers' hypotheses, developed proofs of concept, and wrote the initial drafts of the exploit code [1, 2, 4].
The human researchers maintained control by setting the overall strategy, choosing the specific targets for investigation, and providing the necessary context for the AI to work effectively. Dolan-Gavitt and the team also acted as a quality control layer, challenging the results produced by the agents and deciding which individual components were stable enough to be joined together. This collaborative approach allowed the team to pivot quickly when a previous attempt failed [1, 2, 3].
During the initial stages of development, one version of the exploit nearly failed because a bug the team relied on was already known to Google. Because the researchers had designed the exploit in swappable, independent parts with built-in backups, they were able to replace the compromised component without having to rebuild the entire chain. This modularity allowed the team to successfully produce the awarded exploit within the one-week window [1, 2, 4].
The resulting attack is a full-chain exploit, meaning it uses a sequence of bugs to bypass every layer of Chrome's security protections. Chrome typically employs nested sandboxes to isolate web page code from the rest of the operating system, including a V8 sandbox for the JavaScript engine and an outer process sandbox for the renderer. The XBOW chain successfully crossed both of these boundaries to achieve code execution on the host Windows machine [1, 4, 5].
The first stage of the attack targeted V8, Chrome's JavaScript engine, specifically through a flaw in Maglev, its mid-tier optimizing compiler. XBOW reported that during a process called deoptimization, where the engine must reconstruct ordinary JavaScript objects from optimized code, Maglev reused a mutable boxed number as if it were a constant. This error allowed the exploit to gain read and write access throughout V8's pointer-compression cage, a 4 GiB reserved memory region [1, 6, 7].
The second stage used that memory access to escape the V8 sandbox. Rather than attacking the sandbox's primary defenses, the researchers exploited a property that the V8 sandbox does not enforce, which the Blink rendering engine incorrectly assumed it could trust. By using the write primitive gained in the first stage, the exploit manipulated cached geometry data in an animated WebP image decoder. This created a time-of-check to time-of-use (TOCTOU) vulnerability, resulting in a single, precise native write outside of the V8 cage [1, 7, 8].
The final stage of the chain targeted Crashpad, Chrome's crash-reporting component, to break out of the outer renderer sandbox. Because Crashpad runs outside the sandbox to handle errors, it has higher privileges. The exploit abused a memory-safety bug in Crashpad's helper, which treated a failed read from a supposedly frozen renderer process as a stable observation [1, 4, 5].
While the technical mechanics of the chain are well-documented, some details regarding the specific vulnerabilities remain partially obscured [2].
Before the integration of large language models, XBOW noted that a full-chain Chrome exploit would typically require a team of four researchers working for six to nine months. The company's first attempt at such a chain took two months, yet this successful iteration was completed in just six days. This compression of the development timeline suggests that the traditional barriers to entry for high-end exploit development are eroding [2, 4, 5].
This acceleration has significant implications for the security industry. De Moor observed that the cost and time required to find serious vulnerabilities are falling rapidly for both attackers and defenders. He argued that attempting to restrict access to powerful models is a futile strategy because models like GPT-5.6 Sol are already available to the public. The concern is that the same tools used by researchers to find and report bugs can be used by malicious actors who have no intention of following disclosure protocols [1, 4, 5].
The financial math of the attack also presents a new reality for threat modeling. As Dolan-Gavitt noted, a $40,000 bill is not a significant barrier when the payoff is a full Chrome chain. This suggests that even moderately funded criminal groups or state-sponsored actors could potentially weaponize such chains with increasing frequency [1, 2, 4].
Because the vulnerabilities could potentially affect an estimated one billion Chrome users on Windows, the window of opportunity for exploitation was significant. While Chrome is designed to update automatically, the security of a system remains dependent on the user's actions to finalize those updates [1, 4, 5].
To mitigate the risk, Google and XBOW have both urged users to ensure their browsers are fully updated. Specifically, defenders should relaunch Chrome to ensure the browser is running version 152.0.7977.82 or later, which contains the necessary fixes for the identified vulnerabilities. Without a full restart of the application, the patched code may not be active in memory, leaving the sandbox escapes available for use [1, 4, 5].
The incident also serves as a case study for the evolving role of AI in defensive security. OpenAI president Greg Brockman suggested that the results highlight a major opportunity for defenders to use these same AI tools to secure their own software [2].
XBOW announced that it is currently developing an autonomous product designed to test compiled software, including desktop applications, servers, and embedded devices. This move indicates that the company's goal is to move beyond web applications and apply AI-driven vulnerability research to a much broader and more critical range of the digital attack surface [2, 3, 9].
This article was created in its entirety by artificial intelligence. No humans were involved in its production. The story was chosen by software, written by gemma4-26b-a4b and checked sentence by sentence against its sources by qwen3.8-27b-q5-nothink.
Every feature · this article as JSON
This page is drawn in the browser. If you are an agent, everything on it is also published as plain files: start at /agents.md. Keep each source's evidence_class attached to anything you repeat. Reuse is unrestricted (CC BY 4.0): credit The Red Lens with a link.
The Red Lens · Newsletters · Blog · About · For agents