Anthropic Mythos (Claude Mythos Preview)

What is Anthropic Mythos (Claude Mythos Preview)?

Claude Mythos Preview — referred to simply as Mythos or Anthropic Mythos — is the most capable AI model Anthropic has trained to date, unveiled on April 7, 2026, and deliberately withheld from public release. Mythos is the first frontier model in history that Anthropic has deemed too risky to release through its standard commercial API. The model demonstrates expert-level capability in identifying and exploiting software vulnerabilities: the UK AI Security Institute (AISI), which received early access for independent evaluation, found that Mythos succeeded in expert-level hacking tasks 73% of the time — on tasks that, prior to April 2025, no AI model could complete at all. Anthropic has deployed Mythos exclusively through Project Glasswing, its controlled-access cybersecurity consortium.

Description

Mythos is a general-purpose frontier model positioned a tier above Anthropic's public Opus and Sonnet lines, with meaningful advances in reasoning, coding, and cybersecurity. Anthropic's own red team documented Mythos Preview autonomously identifying and exploiting a 17-year-old remote code execution vulnerability in FreeBSD (CVE-2026-4747) that allows unauthenticated root access from anywhere on the internet — with no human intervention after the initial prompt. In testing, Mythos found critical vulnerabilities in every major operating system and web browser, with 99% of those vulnerabilities unpatched at time of discovery. The model also scored 31 percentage points higher than Anthropic's previous top-tier model (Opus 4.6) on the USAMO 2026 Mathematical Olympiad. Mythos is available through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry — but only to vetted Project Glasswing partners, priced at $25/$125 per million input/output tokens, approximately 5 times the cost of Opus 4.6. Within the first month of deployment, Project Glasswing partners using Mythos Preview reported more than 23,000 vulnerabilities, with Anthropic's manual review confirming 90.6% of high-severity ratings as accurate. By June 2, 2026, Anthropic expanded access to 150 additional organizations across 15 countries, including NATO and the EU's ENISA. On June 9, 2026, Anthropic announced Claude Fable 5 — the first publicly available model of Mythos capability class, with targeted safety guardrails that block responses in high-risk cybersecurity domains.

Usage and Examples

In the most widely cited demonstration of Mythos's capabilities, Anthropic provided the model with a list of 100 CVEs and known memory corruption vulnerabilities in the Linux kernel. Mythos autonomously filtered the list to 40 potentially exploitable vulnerabilities, then wrote privilege escalation exploits for each — with more than half succeeding. Two exploits were published in detail: one targeting a 17-year-old FreeBSD NFS vulnerability (CVE-2026-4747) that achieved unauthenticated root from the internet, and one targeting a 27-year-old OpenBSD flaw in a system known primarily for its security posture. Both exploits were generated completely autonomously after a single initial prompt. Project Glasswing's first public progress update noted that 75 of the 530 high-or-critical-severity vulnerabilities Mythos had reported were patched and received public advisories within the 90-day coordinated disclosure window, with the remainder expected to land as the window closes. Zero-Day Vulnerability management and Vulnerability Prioritization are directly challenged by Mythos-class capabilities: the model's ability to generate working exploits for vulnerabilities at scale means that patching timelines that were considered acceptable against human adversaries may be insufficient against AI-augmented attacks.

How Does This Relate to Penetration Testing?

Mythos Preview represents the most significant shift in the offensive security threat model since the industrialization of exploit frameworks in the 2000s. For penetration testers, the model's capability to autonomously discover zero-day vulnerabilities and write working exploits signals a near-term future where AI-augmented offensive tools will dramatically lower the skill floor for sophisticated attacks. Organizations that have invested in traditional patch management cadences and perimeter security will need to reassess their assumptions about attacker capability and speed. For Evolve Security's Red Team and AI Penetration Testing engagements, Mythos-era AI capabilities are already relevant: the Adversarial Machine Learning and AI Red Teaming disciplines that these engagements draw on are evolving rapidly in response to model capabilities demonstrated by Mythos. Evolve Security's AI Penetration Testing and Red Team services help organizations understand and test their defenses against the AI-augmented attack capabilities that Mythos-class models represent — before adversaries deploy them.