Anthropic internal documents reveal: The new model codenamed "Mythos" has been tested on a small scale, raising safety concerns.

What Happened

AI entrepreneur Matt Shumer discovered that Anthropic’s internal documents were publicly accessible due to a CMS configuration error. The documents revealed that the company is developing a new model codenamed “Claude Mythos,” which is internally positioned as significantly stronger than the current flagship product, Claude Opus 4.6 (known for its reasoning and coding capabilities).

The documents also specifically mentioned cybersecurity risks—the model has the capability to discover and exploit software vulnerabilities.

Why This Matters

• Internal documents corroborate external reports: Mythos belongs to a new “Capybara” internal tier, larger in scale and performance, and has entered early customer testing.
• This aligns with the industry’s trajectory of “larger models leading to better performance” (companies like OpenAI are following a similar path).
• Security issues are brought to the forefront: Anthropic has consistently emphasized explainability and safety boundaries, but the internal documents explicitly label “exploitable vulnerabilities,” which will certainly draw regulatory attention.
• Potential impact on the industry:
  • Enterprise sector: Stronger reasoning and coding capabilities may further apply to R&D, auditing, and analysis work.
  • Competitors: May be forced to accelerate product iterations.
  • Anthropic itself: This leak exposes vulnerabilities in the company’s operational security.

Comparison of Mythos and Opus 4.6

The documents repeatedly emphasize several key points: larger scale, higher performance, already in external testing, with cybersecurity offensive and defensive capabilities. These are not speculations but statements from the leaked documents.

Market Reaction

• Reports indicate that after the news broke, shares in the cybersecurity sector fell.
• Timing is sensitive: It is circulating in the market that Anthropic is preparing for an IPO with an estimated valuation of about $380 billion, making investors particularly sensitive to signals of safety and compliance.

Risks and Opportunities

• On the risk side:
  • The model’s ability to “discover and exploit vulnerabilities” may trigger regulatory scrutiny and negative publicity.
  • The CMS configuration error has exposed the company’s operational security issues.
• On the opportunity side:
  • R&D, auditing, and automated analysis in the enterprise sector may benefit from more powerful models.
  • Competitors face pressure, which may accelerate product iteration across the industry.

Market Impact Assessment

• Importance: High
• Category: Model release, AI safety, AI research

Judgment: Currently in the early stages, information still has marginal pricing effects. The most suitable audience for this event is traders, who can engage in event-driven trading around the cybersecurity sector and model iteration-related assets.