Claude Mythos Preview

Model Overview

Due to its exceptional advanced capabilities, particularly in identifying software vulnerabilities, Anthropic chose not to release it to the general public. Instead, it was launched under "Project Glasswing", a defensive coalition involving major technology partners like AWS, Apple, Google, and Microsoft.## Capabilities

• Advanced Cybersecurity: The model possesses exceptional abilities to identify and exploit software vulnerabilities, often surpassing human experts, and has found thousands of high-severity, zero-day vulnerabilities in major operating systems and web browsers.
• Complex Reasoning: It showcases frontier-level complex reasoning and technical analysis.
• Multimodal Processing: Supports text, image, and code inputs with a massive context window of up to 1M tokens.

Example Use Cases

• Defensive Security Research: Identifying and patching critical flaws in global digital infrastructure and software (under Project Glasswing).
• Life Sciences Research: Advanced biological and healthcare research within highly trusted, restricted access programs.
• Vulnerability Auditing: Conducting deep-dive audits on large codebases to discover zero-day exploits before malicious actors do.

Performance & Benchmarks

• Currently, specific benchmark scores and parameter counts remain undisclosed.
• In practical application tests, the model successfully identified thousands of previously unknown (zero-day) vulnerabilities in major software systems, demonstrating unprecedented performance in cybersecurity domains.

Intended Use & Limitations

Intended Use: This model is strictly intended for defensive cybersecurity research, advanced life sciences research, and enterprise safety testing by vetted partners. It is part of a controlled initiative to strengthen global digital infrastructure.

Limitations & Restrictions:

• Not Publicly Available: The model is highly restricted to vetted partners through Project Glasswing.
• Safety Concerns: Due to its powerful capabilities, there are significant risks of the model being used by malicious actors, which necessitated the restricted release strategy. (Note: A safer, public-facing counterpart with robust safety classifiers was later released as Claude Fable 5).

About Anthropic

Anthropic is an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Known for the Claude family of models, Anthropic focuses on developing highly capable AI models while deeply emphasizing safety, alignment, and responsible deployment strategies, as evidenced by their handling of the Mythos class.