Anthropic just made its most restricted AI model available to every Enterprise customer, without giving anyone direct access to the model itself. That distinction matters, and it tells you something important about where enterprise AI security is heading.

On August 21, Claude Security, Anthropic’s AI-powered vulnerability scanner currently in public beta for Enterprise plan customers, began running on Claude Mythos 5. Until this week, Mythos 5 had no public model identifier, no API endpoint, and no path to access outside Project Glasswing, the joint program Anthropic runs with the US government to give select national security partners access to its most capable cyber AI.

That fence just got a controlled opening.

What Changed

Enterprise customers can now run Mythos 5 directly through Claude Security at claude.ai/security, with no additional add-on or special approval needed. The model connects to a GitHub repository, traces data flows across files, and returns findings categorized by Common Weakness Enumeration (CWE) codes with confidence ratings, severity scores, and suggested patches.

What customers do not get is the ability to freely prompt Mythos 5. The interface is intentionally narrow. You submit a codebase. You receive a defensive artifact: a structured findings report and proposed fixes. The model itself stays behind Anthropic’s infrastructure, and the output is constrained to a defined security workflow.

This is a deliberate architecture choice, not a technical limitation. Anthropic has been explicit that Mythos 5 carries capabilities that could be misused if exposed through an open API. The Claude Security workflow delivers the defensive value, specifically finding and fixing your vulnerabilities, without handing over the offensive potential.

Why This Is a Bigger Deal Than It Looks

Most enterprise security tools use AI as a layer on top of traditional static analysis. They pattern-match against known signatures or run rule-based checks that any experienced attacker has already learned to evade.

Mythos 5 reasons about code the way a senior security researcher does. It understands context. It can follow a data flow from an input field through five layers of processing to a database query and identify that the sanitization in step three is bypassed under a specific condition. That kind of traversal is exactly what finds the vulnerabilities that signature-based tools miss and what attackers actively look for.

Putting that capability inside a structured workflow that any Enterprise customer can run from their browser is a meaningful expansion. Security teams that previously had to either build their own AI pipelines or rely on partners like CrowdStrike, Palo Alto Networks, or Wiz to integrate these capabilities can now run Mythos 5 scans directly.

What This Means for Business

If you run a software team on a Claude Enterprise plan, you now have access to the most advanced vulnerability scanning AI Anthropic has built, tuned to the same capability level previously reserved for government-vetted security programs.

The practical implications:

For security teams: Mythos 5 in Claude Security is a genuine upgrade from what was available before. CWE categorization with confidence and severity ratings gives you a prioritized findings list, not a raw dump of every possible concern. Suggested patches reduce the time between detection and remediation.

For compliance and audit: Automated scan documentation is increasingly expected by cyber insurance providers and enterprise procurement teams. Having a structured, AI-generated findings report from a credible source strengthens your security posture documentation.

For engineering leadership: The GitHub integration means scans can slot into existing workflows without requiring security engineers to manage a separate platform or write custom tooling. You connect the repo, run the scan, get the report.

For enterprise buyers evaluating AI vendors: The way Anthropic has structured this, controlled access to a powerful model through a narrow, defensible workflow, is the pattern that responsible AI deployment looks like at the frontier. It is worth understanding as you evaluate how other vendors handle the same tradeoffs.

The Broader Pattern

This move is part of a consistent strategy from Anthropic. Rather than making its most capable models freely available and hoping users apply them responsibly, the company is building products that channel those capabilities into specific defensible use cases.

Claude Security is one instance. The same logic applies to how Anthropic has handled other sensitive capability areas. The power goes up; the surface area for misuse stays controlled.

For enterprise teams building out their AI security stack, the question is no longer whether AI can find vulnerabilities better than traditional tools. It can. The question is whether you have access to that capability in a form you can actually deploy. Claude Security answers that question with a reasonably clean yes.

Enterprise customers can access Claude Security at claude.ai/security. The beta is free for Enterprise plan subscribers.

---

Enterprise DNA helps business leaders and data teams understand what AI developments like this mean for their operations. Our AI advisory services help organisations build the right AI stack for their specific context.