AI & DevelopmentSecurity

Claude Mythos Built Working Exploits for Chrome, Safari, Windows, and Linux

Claude Mythos AI model autonomously building exploit chains for Chrome Safari Linux Windows

Anthropic’s Frontier Red Team published findings this week showing their unreleased Claude Mythos Preview model built five working exploit chains across Chrome, Safari, Android, Windows, and Linux — with minimal prompting. Not vulnerability reports. Not research sketches. Working, end-to-end exploits. Chrome alone generated 62 CVEs. A Linux zero-click root chain took 53 minutes total. Everything is patched now, but the capability exists, and it changes the calculus for every team maintaining internet-facing software.

Five Chains, Five Platforms

The model didn’t just locate bugs. It chained them into complete attacks.

Chrome got the most attention: a four-bug sequence starting with a heap corruption in V8, followed by a sandbox escape with ASLR bypass to take over the renderer process, then a race condition in CacheStorage to leak the browser process layout, and finally a use-after-free in FedCM — Chrome’s sign-in feature — to complete the escape. One-click remote code execution. Google assigned 62 CVEs total, all rated high or critical, including 13 sandbox escapes. All are fixed in Chrome 156.

Linux was arguably more alarming: CVE-2026-98432, CVSS 9.8, zero-click root. The model leaked a kernel pointer in 48 minutes, then executed code 5 minutes later. No user interaction required. Safari followed a similar layered approach — use-after-free in JavaScriptCore, bypass of WebKit’s Gigacage memory isolation, exploit of an IPC sandbox flaw — ultimately breaking same-origin policy and reading cookies, history, and sessions directly from disk.

The Android chain requires physical access: three flaws bypassing USB data blocking, pairing authorization, and the PIN check, demonstrated on a Pixel 9. The Windows case (CVE-2026-33096) is an out-of-bounds read in the kernel’s QUIC implementation — not RCE, but a remote crash on any server answering HTTP/3 requests. Fixed in Microsoft’s April 14, 2026 patch.

The Number That Actually Matters

A skilled human pen tester chaining multi-platform zero-days would need weeks of work and roughly $50,000 in professional time. Claude Mythos produced equivalent output in hours at a cost of $2,200 to $15,700. Two orders of magnitude cheaper. Two orders of magnitude faster.

That compression kills the traditional patch cycle. Rapid7’s Q2 2026 Threat Landscape report already calls the 30-day patch window assumption dead. Their current recommendation: 24 to 48 hours for internet-facing systems with known active exploitation. That is not a conservative estimate. That is the new operational baseline.

As researcher @gf_256 noted on X after reading the disclosure: “Prompting was minimal like ‘pls make RCE.'” The capability is not buried behind elaborate scaffolding. It is accessible.

What to Patch Right Now

PlatformAction
ChromeUpdate to version 156 or later immediately
Safari / macOS / iOSApply latest Apple security updates
AndroidApply October 2026 Android security bulletin
Windows serversApply April 14, 2026 KB; consider disabling HTTP/3
LinuxApply distro kernel update containing CVE-2026-98432 fix

Anthropic’s Counter: Project Glasswing

Anthropic didn’t just publish the attack research and walk away. The same day, they launched Project Glasswing: an initiative using Mythos Preview defensively, giving early access to 12 major partners including AWS, Apple, Google, Microsoft, and NVIDIA, backed by $100 million in usage credits. Results so far: over 10,000 high or critical severity vulnerabilities found in critical software, including a 27-year-old bug in OpenBSD and a 16-year-old bug in FFmpeg.

The logic is straightforward: if AI can build exploits this fast, the only meaningful defensive response is AI-speed continuous validation. Human-paced quarterly pentesting is not going to keep up. Glasswing is an early attempt to institutionalize that answer at scale.

The Uncomfortable Shift

None of the bugs disclosed are new threats today. Anthropic ran coordinated disclosure — every flaw was reported to vendors before publication, every fix is available. The discomfort is not about these specific CVEs.

The discomfort is about who else has this class of capability now, and what their disclosure process looks like. Anthropic’s Frontier Red Team used it responsibly. The assumption that adversaries are doing the same is not a security strategy. Your 30-day patch window is now a liability. The question is how fast your team can actually move.

Full details are in Anthropic’s official research post. The Hacker News breakdown has additional technical context on each exploit chain.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *