Defenders Get Access to Anthropic's Most Capable Model
Anthropic has announced two significant expansions of its defensive security posture: an upgrade of Claude Security to run on its most capable model, Claude Mythos 5, and the launch of the Defender Advantage Fund (0xDAF) — a $35 million commitment in Claude credits directed at finding and patching vulnerabilities in open source software.
The announcements, published August 21, 2026, build on Project Glasswing — Anthropic's April 2026 initiative that gave a narrow set of vetted defender organizations early access to Mythos 5 before models with similar capabilities reached the general market.
Claude Security: Now Powered by Mythos 5
Claude Security launched in February 2026 as an internal preview under the name Claude Code Security. It entered public beta on April 30, initially running on Claude Opus 4.7. The August update swaps the engine for Mythos 5 — a meaningful capability jump that Anthropic says enables the scanner to trace data flows across files and read Git history rather than relying on static pattern matching.
How It Works
The workflow is deliberately constrained by design:
- An admin enables Claude Security in the Claude admin console for the organization
- Users navigate to
claude.ai/security, select a repository, and trigger a scan - Mythos 5 traces inter-file data flows and inspects Git history to understand how code has evolved
- Each potential finding is subjected to an adversarial self-verification step — the model challenges its own conclusion before surfacing it, intended to reduce false positives
- Findings are returned with CWE category, confidence rating, severity, and a suggested fix
- Users open Claude Code on the web to implement patches; all patches require explicit human approval before merging
Results can be exported to CSV or Markdown and integrated via webhooks into Slack or Jira. Scans can be scoped to a directory and scheduled for recurring runs. Pricing remains token-based at standard Enterprise rates — no separate add-on cost.
Target Vulnerability Classes
Claude Security focuses on vulnerability categories where cross-file reasoning provides the most leverage: memory corruption, injection flaws, authentication bypasses, and cross-file logic errors — bugs that point-in-time pattern matchers frequently miss because the vulnerable condition only becomes visible when tracing data across module boundaries.
The Safety Architecture Tradeoff
The most consequential design decision in Claude Security is what it does not allow: a free-form prompt box. Users interact with structured scan results and patch suggestions, not an open-ended model interface.
Anthropic's stated rationale is that the offensive risk in powerful security AI concentrates in interactive steering. By restricting output to "specific outputs such as a patch for a vulnerability or a security alert," the same Mythos 5 capability can be broadly deployed across Enterprise customers without opening a path for users to steer the model toward generating novel exploits.
This is a meaningful structural constraint, not just a policy. The model that runs Claude Security is the same Mythos 5 that Anthropic's Cyber Verification Program offers to vetted offensive security researchers with reduced safeguards — but the surface exposed to standard Enterprise customers is deliberately narrower.
Defender Advantage Fund: $35M for Open Source
The 0xDAF (Defender Advantage Fund) commits $35 million in Claude credits to organizations working to:
- Find and patch live vulnerabilities in widely used open source software
- Build automated scanning and patching pipelines
- Research defenses against whole classes of attacks rather than individual CVEs
Pilot grants are expected to begin in the coming weeks. Anthropic has not yet named the first recipient organizations, but the fund is positioned as a complement to the Cyber Verification Program — where the CvP provides enhanced model access to vetted researchers, 0xDAF provides compute resources to security-focused non-profits, research institutions, and open source maintainers.
Partner Ecosystem Gets a Mythos 5 Upgrade
Anthropic's enterprise security partners — Wiz, Palo Alto Networks, CrowdStrike, and Accenture — were already running Opus-powered integrations. The August announcement brings Mythos 5 to their workflows.
Wiz reported running an Opus-powered offensive testing agent continuously across 150,000+ production assets per week. Accenture saw security testing coverage climb from approximately 10% to over 80% across 1,600 applications while cutting scan turnaround from 3–5 days to under an hour. Both organizations will benefit from the Mythos 5 capability jump.
What's Coming Next
Anthropic indicated that the Cyber Verification Program — which provides vetted organizations access to Opus and Sonnet with expanded dual-use capabilities for authorized offensive security work — will expand with broader capabilities in coming weeks. Mythos-class access is expected to follow for qualified defenders.
Project Glasswing, which provides Mythos 5 access to U.S. government partners for critical infrastructure organizations, continues in parallel.
For defenders, the pattern Anthropic is establishing is deliberate asymmetry: open source maintainers and vetted security organizations get access ahead of the general market, and the interface they get is constrained toward defensive output. Whether that asymmetry holds as models become more capable — and as adversaries find their own access paths — remains the central tension in AI-powered security.
Sources
- Anthropic — Bringing the Cybersecurity Capabilities of Claude Mythos 5 to More Defenders
- SecurityWeek — Anthropic Expands Mythos 5 Access to More Defenders, Unveils $35M Open Source Fund
- Unite.AI — Anthropic Deploys Claude Mythos 5 in Security Tools, $35M Open-Source Fund
- The New Stack — Anthropic Brings Mythos 5 to Its Claude Security Vulnerability Scanner