Claude Mythos¶

Development line: project:claude-mythos · thread claude-mythos-security-findings
Last event: - · 0 dated since - · Researched: 2026-09-05 · confidence: medium

What it is¶

Anthropic’s invite-only model line for approved cyber defenders and critical-software teams.

• Scans owned codebases for contextual vulnerability findings and proposed fixes.
• Direct access for approved Project Glasswing partners only.
• Powers Claude Security scans without exposing direct model access.

Development line¶

• The dated line is not written up yet; what is known stands in the sections below.

What changed¶

• 2026-04-07 — Anthropic announced Claude Mythos Preview and Project Glasswing. Preview went to launch partners and more than 40 critical-infrastructure organizations for defensive work rather than general release.
• 2026-04-10 — Anthropic’s dated guidance identifies the model as Claude Mythos Preview and frames this step as operational guidance after launch, not a second model release. AI compresses vulnerability-discovery and exploit timelines; Anthropic recommends closing the patch gap and remediating KEV entries urgently.
• 2026-06-09 — Claude Mythos 5 (claude-mythos-5) replaced Preview for eligible Glasswing participants. It shares Fable 5’s base model while removing cyber safeguards for approved users; documented limits were 1M context, 128K output, and $10/$50 per million input/output tokens.
• 2026-08-21 — Claude Security public beta for Enterprise began running repository scans on Mythos 5, returning findings and suggested fixes without direct Mythos prompting.
• 2026-08-31 — Claude Mythos 5.1 was designated a limited-availability Covered Model for approved partners.
• 2026-09-01 — Claude Mythos 5.1 became the current direct-access Mythos model, still invitation-only through Project Glasswing.

How to use this¶

As of 2026-04-10, make no operational change from this line until we retrieve and review the underlying post or source evidence.

1. Enable Claude Security in organization settings under Claude Enterprise. — https://support.claude.com/en/articles/14661296-use-claude-security
2. Connect an authorized GitHub account, then select an owned GitHub.com or GitHub Enterprise Server repository in Claude Security. — https://support.claude.com/en/articles/14661296-use-claude-security
3. Run the scan, then inspect the CWE category, severity, confidence, reproduction path, and suggested fix for each finding. — https://claude.com/blog/bringing-claude-mythos-5-to-more-defenders
4. Open a proposed fix in Claude Code, then have a human review, test, and approve it before implementation. — https://claude.com/blog/bringing-claude-mythos-5-to-more-defenders
5. For direct Mythos 5.1 use, obtain Project Glasswing approval through an Anthropic, AWS, or Google Cloud account team; use model ID claude-mythos-5-1 only after access is granted. — https://platform.claude.com/docs/en/models/mythos-5-1/overview

Best practices¶

• Scan only code the organization owns or has rights to scan. — https://support.claude.com/en/articles/14661296-use-claude-security
• Treat scans as stochastic evidence, not a replacement for validation; severity settings are not configurable. — https://support.claude.com/en/articles/14661296-use-claude-security
• Patch internet-facing systems within 24 hours of an available exploit; remediate KEV entries first and prioritize the remaining queue with EPSS. — https://claude.com/blog/preparing-your-security-program-for-ai-accelerated-offense
• Send a vulnerability report only after human verification; include the vulnerable code path, a runnable reproduction, a proposed patch, and disclose AI involvement. — https://claude.com/blog/preparing-your-security-program-for-ai-accelerated-offense
• Start AI alert triage with read-only access and measure agreement with a human reviewer before expanding automation. — https://claude.com/blog/preparing-your-security-program-for-ai-accelerated-offense

Superseded by this¶

• 2026-06-09 — Claude Mythos Preview is deprecated; eligible users were moved to Claude Mythos 5.
• 2026-09-01 — Claude Mythos 5 is no longer the current direct-access target; eligible partners should migrate to Claude Mythos 5.1.
• 2026-08-21 — Direct prompting is no longer required for Mythos: Claude Security returns bounded defensive findings without exposing the model.

Still unknown¶

• The 2026-04-10 entry has no URL or text, so its exact claim cannot be matched conclusively to the dated Anthropic guidance.
• Current documentation names Mythos 5 for Claude Security scans and Mythos 5.1 for direct partner access; it does not confirm that Claude Security has moved to 5.1.
• No public self-service path to direct Mythos 5.1 access is documented.

Sources¶

Agent brief¶

• Subject: project:claude-mythos, thread claude-mythos-security-findings, 0 dated events - → -.
• Practical note: As of 2026-04-10, make no operational change from this line until the underlying post or source evidence can be retrieved and reviewed.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.