Anthropic Unveils Mythos AI Model for Cybersecurity, Limits Public Access

Anthropic has announced Claude Mythos Preview, a new frontier model positioned above its current Opus offering, but has not made it publicly available. Instead, the model has been distributed to a consortium of 12 companies and approximately 40 critical infrastructure organizations for cybersecurity defense under "Project Glasswing."

Partners in Project Glasswing include Amazon, Apple, Google, Microsoft, Cisco, CrowdStrike, Broadcom, NVIDIA, JPMorgan Chase, Palo Alto Networks, and the Linux Foundation. Anthropic has committed up to $100 million in usage credits for the project and an additional $4 million directly to open-source security organizations.

Dario Amodei, CEO of Anthropic, in a professional setting.

Claude Mythos Preview, internally codenamed "Capybara," represents a new top tier in Anthropic's model hierarchy.

Performance Benchmarks

Mythos Preview demonstrates significant performance improvements over Opus 4.6 across various benchmarks, particularly in software engineering and cybersecurity tasks.

Notably, Mythos Preview achieved a 59.0% score on SWE-bench Multimodal, more than double Opus 4.6's 27.1%.

Cybersecurity Capabilities

Logan Graham, Anthropic's Head of Frontier Red Teaming, described the model's security capabilities as "reckoning." In recent tests, Mythos Preview identified thousands of high-severity zero-day vulnerabilities across major operating systems and mainstream browsers, some of which had existed for decades. This contrasts with Opus 4.6, which previously found around 500 zero-day vulnerabilities in open-source software.

Digital screen showing lines of code with red highlights, symbolizing discovered software vulnerabilities.

Beyond vulnerability discovery, Mythos Preview can autonomously generate exploit code. Anthropic's technical blog provided several examples:

• OpenBSD Vulnerability: The model discovered a 27-year-old integer overflow vulnerability in the TCP protocol's SACK mechanism within OpenBSD, an operating system known for its security. This flaw could allow an attacker to crash a target machine by connecting to it.

• FFmpeg Vulnerability: Mythos identified a 16-year-old vulnerability in FFmpeg, a widely used video dependency library, which had gone undetected despite 5 million automated scans.

• Linux Kernel Privilege Escalation: The model autonomously chained multiple vulnerabilities to achieve a privilege escalation attack, gaining full control over a machine from ordinary user permissions.

In an interview, Graham detailed Mythos Preview's capabilities, which include advanced exploit development, penetration testing, endpoint security assessment, system misconfiguration troubleshooting, and direct analysis of software binary files without source code. The ability to analyze binaries is particularly significant, as real-world attackers typically lack access to source code.

Dario Amodei, Anthropic's CEO, stated in a release video that while the model was not specifically trained for cybersecurity, its proficiency in code led to its strong security capabilities. He also noted that more powerful models are expected from Anthropic and other companies, necessitating new solutions.

Previously, Opus 4.6 had a near-zero success rate in autonomously developing exploit code. In contrast, Mythos Preview succeeded 181 times in a Firefox JavaScript engine exploit test, compared to Opus 4.6's two successes out of hundreds of attempts. Anthropic engineers, without security backgrounds, were able to generate working remote code execution exploits overnight using Mythos Preview with simple instructions. Graham anticipates that models from other AI companies may achieve similar capabilities within 6 to 18 months.

Abstract image of a digital lock being broken by glowing data streams, symbolizing exploit generation.

Public Availability and Cost

Claude Mythos Preview currently lacks a public API, announced pricing, or a confirmed public release date. While it is not listed in Anthropic's API documentation, Google Cloud has made Mythos Preview available to select customers for Private Preview access on Vertex AI.

Graham likened this phased release to the cybersecurity industry's "coordinated vulnerability disclosure" practice, allowing software developers time to patch flaws before public disclosure. This strategy suggests Anthropic is providing partners like Amazon, Apple, Google, and Microsoft an opportunity to use Mythos Preview to scan and patch their systems before a broader public release.

Internal Anthropic documents indicate the model is "very expensive to run" and will be costly for customers. While Opus 4.6 is priced at $5/$25 per million input/output tokens, Mythos Preview's Glasswing partnership pricing is $25/$125, five times higher. This pricing is reportedly slightly cheaper than GPT-5.4 Pro.

Industry analysis suggests a realistic timeline for expanded early access in mid-2026, with a public API release in Q4 2026. This timeline may align with Anthropic's reported plans for an IPO around October 2026. Anthropic's blog also states a goal to "enable users to safely deploy Mythos-level models at scale" for general scenarios beyond cybersecurity.

Context and Background

The announcement coincides with Anthropic's projected annualized revenue for 2026 exceeding $30 billion, more than tripling the previous year's $9 billion. On the same day, Anthropic also announced computing power agreements with Google and Broadcom, securing approximately 3.5 gigawatts of Google AI processor computing power. Bloomberg also reported that Anthropic hired Eric Boyd from Microsoft to lead infrastructure expansion.

Vast, futuristic data center with glowing server racks, symbolizing immense computing power.

The existence of Mythos was inadvertently revealed last month when a misconfigured content management system exposed approximately 3,000 internal documents, including a draft blog post about the model, in a public data store. Anthropic attributed this to human error.

Anthropic is currently involved in a lawsuit with the Pentagon, which earlier this year listed the company as a "supply chain risk" due to restrictions on technology use. A federal judge temporarily blocked the enforcement of this designation. Anthropic has stated it has been communicating with the U.S. Cybersecurity and Infrastructure Security Agency (CISA) and the Department of Commerce regarding Mythos Preview.

Industry Implications

While Anthropic's simultaneous development of powerful AI systems and warnings about their dangers may appear contradictory, the reported cybersecurity capabilities of Mythos are significant. Independent evaluations from partners, including CrowdStrike CTO Elia Zaitsev and Microsoft, corroborate Anthropic's claims of improved vulnerability discovery and exploitation.

Graham stated that if Project Glasswing remains limited to a few companies, it would be considered a failure, indicating Anthropic's ambition to drive a broader industry shift in security practices.

For general developers and users, access to Mythos-level models is not expected in the short term. The timeline for public release may depend on competitive pressures. Graham suggested that other AI companies might achieve similar capabilities within 6 to 18 months. Should competitors like OpenAI's GPT-6, Google's Gemini 4, or DeepSeek's V4 match or surpass Mythos's capabilities, Anthropic may be compelled to accelerate its public release.