XBOW’s security team linked three previously unknown Chrome flaws into a full exploit that runs code on a Windows machine when a malicious page loads (The Hacker News). They completed the attack in a six‑day sprint by directing AI agents running OpenAI’s publicly available GPT-5.6 Sol model. Google has verified all three bugs, issued patches, and paid XBOW the maximum $250,000 bounty.
The chain breaks through every protection layer in Chrome, moving from a standard web page through nested sandboxes straight to the host OS. It works without any user clicks or downloads, relying solely on page loading. XBOW says the bugs were never used against real users and have since been patched.
The six‑day build window ran from Sunday, August 2, to Friday, August 7, coinciding with the Black Hat and DEF CON conferences in Las Vegas. Months of prior work set the stage: an internal challenge launched in January and a preliminary full Chrome chain completed over two months. XBOW’s logs show the AI run cost about $40,000 in usage fees.
Human researchers and autonomous AI agents split the workload. The agents searched Chrome’s source, tested ideas, generated proofs of concept, and drafted initial exploit code. Humans defined targets, supplied context, evaluated outputs, and stitched the pieces together.
Researcher Brendan Dolan‑Gavitt, who worked alongside the agents, highlighted the speed gain. He noted that before large language models, building a full Chrome chain typically required a four‑person team for six to nine months; their first AI‑assisted attempt took two months, and the awarded submission was finished in six days.
The three vulnerabilities hit different browser components. The first flaw, in V8—the JavaScript engine—carries CVE‑2026‑85045. The second bug affects the animated WebP image decoder, and the third targets Crashpad, Chrome’s crash handler, under CVE‑2026‑78952. Google’s release notes credit Dolan‑Gavitt with discovering two of the issues.
The JavaScript engine bug gave memory control, which the animated WebP decoder leveraged to escape the inner sandbox. Google patched the Crashpad issue in version 152.0.7977.65 on August 25 and fixed the V8 bug in version 152.0.7977.82 on September 3. Although Chrome updates automatically, security teams recommend restarting browsers to ensure the new versions are fully active.
The rapid development of a full-chain browser exploit using publicly available AI models highlights a shifting dynamic in software security and vulnerability research. By leveraging automated agents for code analysis and proof-of-concept generation, researchers compressed a timeline that traditionally spans months into a matter of days. This acceleration suggests that threat actors may increasingly adopt similar automated workflows to discover and weaponize zero-day vulnerabilities faster than ever before, raising the urgency for automated defensive patching and proactive code hardening.
Spotted an error? See our corrections policy or email editorial@lordofclaude.com.
Discussion & Comments (0)
Join the conversation