Claude Mythos Preview: The AI Security Breakthrough of 2026

AI-Security

2026-04-08

Claude Mythos Preview autonomously finds zero-days and writes working exploits - 1M token context, April 2026 launch. What it means for security teams.

Frequently Asked Questions

Claude Mythos Preview** is Anthropic's most capable frontier model, designed for autonomous cybersecurity research and agentic coding. Launched April 7, 2026, it is available only through Amazon Bedrock (US East) to an allowlisted set of internet-critical organizations and open-source maintainers. General public access is not planned in the near term due to dual-use risk. For background on how AI capabilities are evolving, see the [AI Meets Blockchain overview](https://www.ancilar.com/knowledge-hub/blogs/ai-meets-blockchain).

**Project Glasswing** is Anthropic's coordinated initiative to use Mythos proactively to harden critical software before these capabilities proliferate to potential adversaries. It partners with selected organizations, prioritizing internet-critical infrastructure, to discover and patch vulnerabilities first. Anthropic describes it as a "starting point" for broader cross-institutional collaboration on **AI cybersecurity model** deployment. [Get in touch with Ancilar](https://www.ancilar.com/contact) if you want to explore what proactive AI security review looks like for your team.

Yes. While Mythos's autonomous exploit-writing capability is far beyond what current public models offer, **Claude Code plus Sonnet** is already a practical tool for AI-assisted security review, threat modeling, and vulnerability pattern detection in everyday development workflows. Teams that are not running AI-assisted code review are already behind the curve. See [how AI-assisted auditing works in practice](https://www.ancilar.com/knowledge-hub/blogs/ai-smart-contract-auditing) as a starting point.

Releasing a model capable of autonomously writing working exploits for major OS vulnerabilities would create unacceptable risk of enabling sophisticated **zero-day vulnerability AI** attacks at scale. Anthropic's November 2025 disclosure of the first AI-orchestrated cyber espionage campaign, using Opus, confirmed that agentic AI capabilities are already being weaponized. Broader access follows only after Glasswing's hardening work creates a safer foundation.

Copybara is Anthropic's internal designation for the 4th tier of Claude models, above Haiku, Sonnet, and Opus. Claude Mythos Preview is the first model released at this tier. It is a qualitative step change: deeper reasoning, a 1M token context window, and autonomous task execution that goes well beyond what Opus can do.