# Claude Mythos in Cybersecurity: Autonomous Zero-Day Discovery, CVE-2026-4747 | Alher Tech

> How Claude Mythos autonomously discovered thousands of zero-day vulnerabilities including CVE-2026-4747 (FreeBSD NFS RCE) and a 27-year-old OpenBSD TCP SACK bug.

- Canonical page: https://alhertech.com/en/claude-mythos/cybersecurity/
- Site: Alher Tech (custom software, AI agents and SEO engineering, https://alhertech.com/)
- Contact: https://alhertech.com/en/contact/

---

In the weeks before its April 2026 announcement, Anthropic ran Claude Mythos against every major operating system, browser and critical infrastructure stack. The model autonomously surfaced thousands of zero-day vulnerabilities, including a 27-year-old OpenBSD TCP SACK remote code execution bug and a 17-year-old FreeBSD NFS RCE now tracked as CVE-2026-4747. This is the page that explains how, why it matters, and the policy decisions that followed.

Updated: May 9, 2026

## What Mythos actually does in a security audit

Given a target codebase and a vulnerability class hint (e.g. "check for memory safety bugs in the TCP stack"), Mythos clones the repo, builds it, fuzzes the surface, reads syscall paths, traces data flow, drafts a candidate exploit, runs it under a sandbox and writes the patch. The whole loop takes 10–90 minutes per target, orders of magnitude faster than a human researcher and with comparable patch quality on a sample of disclosures Anthropic released.

## CVE-2026-4747: FreeBSD NFS RCE

CVE-2026-4747 is a remote code execution flaw in the FreeBSD NFS server's request parsing path that survived 17 years of public review. Mythos identified the bug in roughly 25 minutes of unsupervised analysis. Anthropic disclosed it through a 135-day private channel with the FreeBSD Foundation; the patched release shipped before the public announcement. This is the first major CVE assigned where the discoverer is not a human researcher but a frontier model run autonomously.

## The 27-year-old OpenBSD TCP SACK bug

Selective Acknowledgment (SACK) handling in OpenBSD's TCP stack contained a remote code execution flaw that dated back to 1998. Mythos surfaced it during a routine kernel sweep. The bug had been touched by ~14 commits over its lifetime, none of which addressed the underlying state machine misuse. The disclosure landed with a CVE issued and the OpenBSD team publicly noting that no human had spotted the bug in nearly three decades of audit cycles.

## Why this triggered Project Glasswing

The same capability that lets Mythos discover and patch zero-days at scale would, in adversarial hands, let an attacker mass-produce them. Anthropic's Responsible Scaling Policy treats autonomous offensive cyber capability as a high-risk capability requiring controlled deployment. The result is Project Glasswing: invitation-only access for 12 founding partners and around 40 vetted critical-infrastructure operators, with a 135-day vulnerability disclosure window before any finding goes public.

## How defenders use Mythos in production

Glasswing partners feed their own private codebases into Mythos under NDA. The agent runs against the patched-and-deployed branches every release, surfacing regressions and dormant vulnerabilities that would have shipped to production. Microsoft, Apple, Google and Amazon are confirmed launch partners; the program effectively gives the West a head start on a class of vulnerability discovery that adversaries will eventually replicate.

## Frequently asked questions

### Is Claude Mythos used offensively?

Anthropic states no: every Glasswing engagement is defensive, with disclosure timelines and operator vetting. AISI's published evaluation independently confirmed the model has the capability to chain exploits autonomously, which is precisely why deployment is restricted.

### Can my company get access for security audits?

Glasswing is currently capped at ~40 organisations. Anthropic prioritises critical-infrastructure operators (telecoms, energy, financial market infrastructure, healthcare networks) and major OS / browser vendors. Other organisations should wait for the eventual public release or partner with a Glasswing member.