LLM Wiki

Appearance

Built 26/04/17 09:39commit 8de3d61

Claude Mythos Preview System Card ​

中文 | English

Summary ​

This source is Anthropic's 245-page system card for Claude Mythos Preview. It combines capability benchmarking, Responsible Scaling Policy analysis, cybersecurity testing, alignment assessment, model-welfare analysis, and release-governance rationale into one artifact, with the most important operational conclusion being that Anthropic chose not to make the model generally available.

Source ​

Key Contributions ​

  • Provides a single canonical source for Anthropic's own description of Claude Mythos Preview's capabilities, safeguards, evaluation methods, and release decision.
  • Makes the release posture explicit: Anthropic says the capability jump is large enough that the model is being limited to a defensive cybersecurity program with selected partners instead of general availability.
  • Bundles several otherwise separate evaluation surfaces into one artifact: RSP and Frontier Compliance work, cyber evaluations, alignment assessment, model-welfare work, benchmark results, and qualitative impressions.
  • Extends the public Mythos Preview discussion beyond cybersecurity alone by including welfare, hallucination, constitution-adherence, and safeguard-evasion analysis.
  • Acts as a durable primary source for later claims about Mythos Preview, because it records both the benchmark picture and the safety-governance framing in one place.

Strongest Claims ​

  • Claude Mythos Preview is Anthropic's most capable frontier model to date and shows a large step up over Claude Opus 4.6 on many benchmarks.
  • The capability increase is significant enough that Anthropic decided not to release the model generally at this stage.
  • Anthropic treats Mythos Preview as strong enough to justify extensive cyber, alignment, welfare, and safeguard-evasion review before broader deployment decisions.
  • The system card positions current findings as inputs into future Claude releases and future safeguard design, not as a one-off report about a single model.

Scope Notes ​

  • This source is broader than the existing red.anthropic.com Mythos Preview cybersecurity post in raw/anthropic/claude-mythos-preview/cybersecurity-assessment.md; the blog post focuses on cyber capability case studies, while this system card is the fuller release-and-evaluation record.
  • Because the raw source is a PDF converted into markdown, some tables, figure references, and page-layout details may read less cleanly in raw/ than in the original PDF.

Visual Notes ​

  • The maintained raw markdown keeps local figure evidence inline: raw/anthropic/claude-mythos-preview/system-card.md currently embeds 86 rendered page images from raw/anthropic/claude-mythos-preview/assets/system-card/pages/.
  • Those rendered pages matter most where the PDF carries dense tables, benchmark charts, or qualitative figure layouts that are easier to verify visually than from flattened markdown alone.
  • For later patrols, this means the markdown raw file is the main reading surface, but the local page renders remain available when a claim depends on the exact figure, table, or page composition.