Microsoft Tests Claude Mythos to Mitigate Vulnerabilities

Microsoft is testing Anthropic's Claude Mythos Preview under a cooperative initiative called Project Glasswing, teaming with partners including Amazon, Apple, CrowdStrike, and Palo Alto Networks to identify and mitigate vulnerabilities earlier and coordinate defensive responses. Microsoft says evaluations using its open-source detection benchmark produced substantial improvements relative to prior models, and it plans to fold Claude Mythos Preview and other advanced models into its SDL to surface issues and provide mitigation guidance.

Microsoft evaluated Claude Mythos Preview against a real-world detection engineering benchmark and reports meaningful performance gains over earlier models. Partners in Project Glasswing gain early access to the preview to run adversarial and defensive workflows, threat modeling, and vulnerability discovery. Key integration points Microsoft calls out for practitioners include:

• •Early-stage vulnerability discovery embedded in the SDL
• •Automated detection engineering and triage powered by Claude Mythos Preview
• •Playbook generation and customer guidance to reduce AI-driven exposure

This is a defensive-first application of a frontier LLM, not a consumer product launch. Microsoft coordinating with Anthropic and major infrastructure and security vendors signals two trends: first, leading organizations treat large models as tools for both offense and defense; second, multi-vendor red-team/blue-team testing is becoming standard practice for model rollout in security-critical contexts. The project also reflects an industry move to operationalize models inside security lifecycles rather than treating them as external services.

Quote and posture "Security is a team sport," Holecek wrote, underscoring Microsoft's collaborative stance. Microsoft also urged customers to stay current on patches and reduce exposure, adding "The time to act is now and we look forward to partnering with the industry to build a safer world for all."

Look for published benchmark artifacts, integration patterns for SDL pipelines, and responsible-disclosure workflows from Project Glasswing participants. The specific scope of model access, threat scenarios tested, and measurable detection/false-positive trade-offs will determine how broadly defenders adopt model-driven workflows.