Anthropic Routes Mythos 5 Into Defensive Security Tools

Anthropic is deploying Claude Mythos 5 to enterprise security teams, cybersecurity vendors, and open-source maintainers through defensive tools rather than unrestricted model access. The August 21, 2026 rollout expands access to the model’s cyber capabilities while keeping exploit-generation pathways outside the user interface.

Claude Security Runs on Mythos 5

The central product update is Claude Security, which now uses Claude Mythos 5 for Claude Enterprise customers. Administrators enable the public-beta scanner from the admin console and connect repositories through Anthropic’s GitHub App.

Mythos 5 analyzes data flows across files and examines Git history, giving it a broader view than traditional pattern-matching scanners. Each result passes through an adversarial verification step in which the model challenges its own finding before returning a structured report.

Reports include a Common Weakness Enumeration (CWE) category, severity rating, confidence score, and suggested remediation code. Developers can open Claude Code on the Web at claude.ai/code to review and apply proposed fixes, while human sign-off remains required for every patch.

This interface design is a security boundary. Users receive findings and defensive patches instead of an open prompt box that could be redirected toward exploit development. Teams evaluating the workflow can also compare it with Anthropic’s earlier Claude Security beta, which used Opus 4.7 before the Mythos 5 upgrade.

Enterprise Pricing and Workflow Costs

Claude Security uses existing Enterprise Extra Usage billing, with no separate software add-on fee. Scans are charged as token consumption at $10 per million input tokens and $50 per million output tokens.

Workflow component	Availability or cost
Claude Security engine	Claude Mythos 5 for Claude Enterprise customers
Repository connection	Anthropic GitHub App
Finding output	CWE, severity, confidence, and remediation code
Patch application	Human review and sign-off required
Input usage	$10 per million tokens
Output usage	$50 per million tokens

For engineering organizations, the cost model shifts attention toward repository scope and scan frequency. Large monorepos can generate substantial input-token usage, especially when cross-file analysis and Git history are enabled. Establishing repository boundaries and scheduling scans around meaningful code changes will matter more than treating the scanner as a free continuous service.

Security Vendor Integrations

Anthropic is also working with cybersecurity vendors and service providers, including Palo Alto Networks Unit 42. Integrations target security operations, incident response, threat intelligence, and detection engineering.

Partner systems use structured APIs to return exposure analyses, security alerts, and remediation code. The API approach preserves the same separation between defensive outputs and unrestricted cyber interaction, giving vendors a way to embed Mythos 5 into existing analyst workflows.

Teams building their own agentic security workflows should apply the same principle. Agent security guidance becomes especially relevant when a model can inspect repositories, reason across many files, or propose changes to production code.

$35 Million for Open-Source Remediation

Anthropic launched the Defender Advantage Fund (0xDAF) with $35 million in Claude credits for open-source infrastructure. The program focuses on patching live vulnerabilities in widely used projects, automating scanning and patching loops for maintainers, and funding architectural research into broader classes of attacks.

The funding addresses a capacity problem created by faster vulnerability discovery. During Project Glasswing, Mythos identified more than 10,000 critical flaws in a month, exceeding the ability of human maintainers to triage and repair them. Discovery without remediation increases operational pressure, even when the findings are accurate.

The timing also aligns with European Union Cyber Resilience Act vulnerability-reporting mandates taking effect on September 11, 2026. Open-source projects that become part of regulated software supply chains will need processes capable of handling both disclosure volume and repair verification.

Expanded Access for Vetted Researchers

The Cyber Verification Program (CVP) will expand over the coming weeks to cover broader dual-use capabilities on Claude Opus and Claude Sonnet. Anthropic plans to follow with safeguarded Mythos-class access for authorized defenders.

For developers, the practical boundary is clear: use Claude Security or an approved partner integration for repository analysis, and route higher-risk research through a vetted program rather than a general-purpose chat surface. Treat every generated patch as a candidate change, validate it with tests and independent review, and budget token usage against repository size before enabling recurring scans.