Q: What is Claude Mythos?

What is Claude Mythos?

Here's a comprehensive overview of **Claude Mythos**: ## What is Claude Mythos? **Claude Mythos** (specifically **Claude Mythos Preview**) is Anthropic's newest and most powerful large language model, announced in April 2026. It represents a "step change" in AI capabilities compared to previous models like Claude Opus 4.6 and Sonnet 4.6. ### Key Capabilities - **Cybersecurity powerhouse**: Mythos can autonomously find and exploit zero-day vulnerabilities in major operating systems and web browsers. In internal benchmarks, it achieved full control flow hijack (tier 5) on ten fully patched targets — a dramatic improvement over prior models. - **Autonomous exploit development**: Non-experts at Anthropic used the model to produce working remote code execution exploits overnight without formal security training. - **Multi-vulnerability chaining**: The model chained four vulnerabilities together to escape both renderer and OS sandboxes in a web browser exploit. - **Massive context window**: Supports **1 million tokens** of context and **128,000 tokens** of output. - **Adaptive reasoning**: Can adjust its reasoning depth based on task complexity. ### Technical Specifications | Feature | Detail | |---|---| | Context window | 1M tokens | | Max output | 128K tokens | | Availability | Gated research preview | | Cloud platforms | Amazon Bedrock (us-east-1), Google Cloud Vertex AI | | Endpoint type | bedrock-mantle (not standard bedrock-runtime) | | Prompt caching | Supported (min 4,096 tokens per checkpoint, max 4 checkpoints) | ### Project Glasswing Alongside Mythos, Anthropic launched **Project Glasswing** — a defensive initiative to use the model to help secure critical software and prepare the industry for AI-driven cyberattacks. Anthropic notes that over 99% of vulnerabilities found by Mythos Preview have not yet been patched, limiting public disclosure. ### Important Context Anthropic emphasizes that these cybersecurity capabilities **emerged as a downstream consequence of general improvements in code, reasoning, and autonomy** — not from explicit security training. The company believes that in the long term, such models will benefit defenders more than attackers, but acknowledges the transitional period may be "tumultuous." The model is currently available only as a **gated research preview** to select partners and researchers.