Skip to content
trending

News

Anthropic Mythos Found Rejetto HFS Flaw, Attacks Within a Day

Anthropic's Mythos model discovered a critical flaw in Rejetto HFS file server. Attackers exploited it within 24 hours of disclosure.

By 3 min read
Anthropic Mythos Found Rejetto HFS Flaw, Attacks Within a Day

Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.

👁 114 reads

Anthropic’s Mythos model discovered a critical vulnerability in the popular Rejetto HFS file server, and attackers exploited it within a day of disclosure. This is only the second of 286 vulnerabilities found by Mythos known to have been used in real-world attacks.

The case was reported on October 3, 2026 by The Register. The vulnerability, designated CVE-2026-61500, was found by Zach Hanley of Horizon3, which has had access to Mythos since July 2026 as part of Project Glasswing. Anthropic only provides this model to selected security partners because it considers it too powerful for general release.

How the Vulnerability Works

Rejetto HFS is a free HTTP file server often used for quick file sharing over networks. In versions 3.0.0 through 3.2.0, the key used to sign session cookies was generated using JavaScript’s Math.random(). The same random number generator provided values to unauthenticated users during login. According to Horizon3’s description, Mythos:

  • noticed that the xorshift128+ algorithm behind the generator is fully reversible;
  • proposed using Microsoft’s Z3 solver to reconstruct the generator’s seed from leaks and wrote the constraints for the solver;
  • combined this with the fact that HFS’s administrative API allows adding custom endpoints that execute JavaScript;
  • built a working exploit: forged administrator session, then remote code execution.

Hanley emphasized that Mythos didn’t simply flag a “dangerous generator” but independently connected two weak points into a complete attack chain.

From Disclosure to Attack: About 24 Hours

  1. September 30 (Wednesday) - Horizon3 publishes vulnerability details.
  2. October 1 (Thursday) evening - Patrick Garrity of VulnCheck reports that the company’s honeypots detected the first attacks from a single IP address in China, targeting servers in the US and Japan.
  3. October 2 (Friday) - four more attempts from two US addresses in the same subnet, most likely through a proxy.

According to a tally maintained by Garrity, Mythos and Project Glasswing have 286 CVE numbers to their credit, and as of Thursday only one of those vulnerabilities had been exploited in real attacks. The fix is in HFS version 3.2.1, which also patches a second vulnerability, CVE-2026-61501.

Not everyone is impressed. The Register’s readers point out that reconstructing Math.random() state using Z3 is a technique known for years. Mozilla’s documentation has long warned that this function is unsuitable for cryptography. The discovery is therefore not revolutionary in theory, only in speed: the model independently assembled known elements into a ready-made attack.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

What This Means for You

  • If you’re running Rejetto HFS version 3.x anywhere, update it to 3.2.1 or cut off internet access today. HFS was previously on CISA’s list of actively exploited vulnerabilities (CVE-2024-23692).
  • The time between vulnerability disclosure and attacks is shrinking to hours. Models like Mythos accelerate defense, but the same techniques quickly reach attackers.

You can read more about how banks and companies are responding to models like Mythos in our article on AI in bank cybersecurity. For more about the company itself, see our guide to Anthropic, and for publicly available models, check our guide to Claude.

Sources: The Register (3.10.2026), Horizon3.ai, posts by Patrick Garrity (VulnCheck). Verified October 4, 2026.

Kinetyra

// read next

More on this from QUANTSLANT

Piotr Olszewski

Editor and publisher · Quantslant

I edit and publish Quantslant. AI tools help with research, drafts and translation; choosing topics, running the tests, checking facts and the decision to publish are mine, and so is the responsibility for what appears here.

read nextMeta AI in WhatsApp: What It Sees and How to Limit ItguidesMeta AI in WhatsApp: What It Sees and How to Limit ItMeta AI in WhatsApp appears as a colorful circle on your Chats tab. It doesn't read your private messages, but you can block it in specific chats.

Most read

  1. 1Best AI Video Generator Apps for Your Phone (2026)
  2. 2Claude Opus 5.5 Motion Graphics: 60s Showreel from One Prompt
  3. 3Wan 2.2 on RTX 4090: timings, VRAM and four test clips
// site map

This page in other languages1

Quantslant in numbers 1,427posts in archive26news from the last 7 days18video models in the observatory
🎬 AI Video Benchmarks: Higgsfield DoP vs OpenAI Sora • Higgsfield vs Runway Gen-3 • AI Video Credit Calculator
× ‹ enlarge ›