Anthropic has announced it is expanding access to its tightly controlled Mythos 5 AI cybersecurity model. Through security product integrations, an updated Claude Security service, new funding for open-source projects, and a broader verification program for authorized users, the company is positioning the rollout around controlled delivery rather than direct access.
Mythos 5 will operate in the background of partner tools used for security operations, incident response, detection, and code security, with customers receiving defined results such as vulnerability findings, patch recommendations, and alerts.
The expansion also includes a $35 million pool of Claude credits through what is known as the Defender Advantage Fund, which is designed to help organizations supporting open-source security. Taken together, the announcements show how Anthropic is extending the reach of its most capable cyber technology beyond its initial tightly controlled program of users.
More Routes to Use the Model, but Fewer Ways to Access It Directly
The move builds on Project Glasswing, Anthropic’s earlier program that provided a limited group of organizations with access to its highest-tier cyber AI model, Mythos. The stated objective was to help defenders identify and resolve software weaknesses before similar capabilities became broadly available.
Mythos 5, the successor to Mythos, is now being introduced more broadly through cybersecurity partners working across security operations, incident response, and detection. It is designed to operate through purpose-built interfaces, with the system intended to return only the output required for a defensive task. This allows users of these partner platforms to take advantage of the model’s strengths without giving them full access. Anthropic said this is intended to help organizations operating in high-consequence sectors, including healthcare, utilities, financial services, and software supply chains.
Claude Security, which remains in public beta for Claude Enterprise customers, is also receiving an upgrade. Its codebase scanning capabilities will run on Mythos 5, producing findings categorized by common weakness type, severity, and confidence, alongside proposed remediation.
The changes do not remove people from the software development process. Suggested fixes must still be taken through Claude Code and approved by a human before deployment, preserving a checkpoint between AI-generated analysis and changes to production systems.
Partners May Hold the Keys to Mythos 5
Anthropic’s approach to expanding access to its model is one that could determine business procurement strategy when it comes to security vendors. If the most capable cyber models are not being given directly to customers, a vendor’s AI partnerships may become as important as its detection, response, and scanning features.
