Anthropic’s Mythos 5 Upgrades Its Claude Security Scanner

Anthropic has put its most capable cyber model inside a product enterprises can switch on themselves. As of August 21, Claude Security, the company’s codebase vulnerability scanner, runs its scans on Claude Mythos 5. This is the model Anthropic kept away from general release because of its offensive security capabilities. Alongside the upgrade, Anthropic launched a $35 million Defender Advantage Fund (0xDAF) for open-source security and outlined plans to widen access through partners and its Cyber Verification Program.

From restricted model to enterprise scanner

Mythos 5 arrived in June alongside Fable 5, which shares the same underlying model but ships with strict guardrails that block dual-use cyber work. Until now, Mythos 5 reached only vetted organizations through Project Glasswing, roughly 150 partners according to The New Stack.

Anthropic’s case for widening access rests on the difference between access to a model and access to its outputs. The riskiest scenario is a user prompting the model directly and steering it toward harm. If users receive only a defined artifact, such as a finding or a patch, Anthropic argues that the risk drops sharply.

How the scan works

Claude Security is in public beta for Claude Enterprise customers. Admins enable it in the admin console, and users pick a repository to scan. Each finding comes back with a CWE category, confidence and severity ratings, and a suggested fix. According to MarkTechPost and Dataconomy, the scan traces data flows across files and reviews Git history rather than relying on pattern matching, with an adversarial verification pass to cut false positives.

Fixes are applied in Claude Code on the web using the models the organization already has, and every patch needs human review and approval. Scans are billed as standard token usage with no add-on, though Mythos 5 is not cheap at $10 per million input tokens and $50 per million output tokens.

The same “outputs, not access” design will extend to partner products. Anthropic is working with security vendors to embed Mythos 5 in detection, incident response, and remediation tools. End users there receive results through a purpose-built interface and cannot prompt the model to, for example, write an exploit.

The real bottleneck is patching, not finding

The fund targets a problem Glasswing itself exposed. Partners uncovered more than 10,000 high- or critical-severity vulnerabilities in the program’s first month. By Anthropic’s own count at the time, only 75 of the 530 high- and critical-severity bugs disclosed to open-source maintainers had been patched, and some maintainers asked Anthropic to slow its disclosures.

0xDAF puts $35 million in Claude credits toward three areas: patching live vulnerabilities in widely used projects, automating scanning and patching in reusable ways, and pursuing approaches that resist entire classes of attack. It builds on $4 million in direct donations made through Glasswing. The Next Web raised a fair limit: credits are not maintainers, so the fund mainly helps projects that already have people to put them to work.

Open questions for DevSecOps teams

The New Stack flagged a gray area in the “scan code you own” framing. Most enterprise repos include open-source libraries. Scanning a vendored copy of a popular package would surface the same flaws an attacker wants, and those flaws exist everywhere that library ships.

For teams evaluating the beta:

• Plan remediation capacity first. A scanner that finds more bugs than your team can fix just grows the backlog.
• Route findings into existing workflows, such as ticketing, code review, and change management, rather than into a separate queue.
• Decide how vendored and third-party code is handled, including disclosure to upstream maintainers.
• Budget token costs per repository. Frontier-model scanning at scale adds up quickly.
• Keep humans on approvals. Anthropic requires it, and internal policy should too.

Vulnerability discovery is becoming cheap and fast. The organizations that benefit will be the ones that can triage, patch, and ship fixes at the same pace.