Claude Security scans now run on Mythos 5 for Enterprise

Anthropic is putting Claude Mythos 5 to work for Enterprise security teams — without handing them a prompt box on the model.

As of August 21, 2026, Claude Security scans run on Mythos 5 in public beta for Claude Enterprise customers. Point the product at a GitHub repository and the model traces data across files, reasons about how components interact, and returns findings with a CWE category, confidence and severity ratings, and a suggested fix. Suggested patches then open in Claude Code on the web, using the models the team already has. Mythos stays behind the scan. It does not follow the user onto other surfaces, and nothing lands in the repo until a human reviews and approves it.

That packaging is the story. Until now, Mythos-class vulnerability discovery lived inside Project Glasswing, the April program that put Claude Mythos Preview and later Mythos 5 in front of a small group of organizations securing critical software. This is the first self-serve path to those defensive results. Scans bill as ordinary token usage under the existing Enterprise plan. There is no separate Mythos add-on.

Findings, not a chat window

Anthropic’s stated risk model is simple: the dangerous case is direct access, where someone can steer a frontier cyber model toward an exploit. Claude Fable 5 was the first containment step — the same class of model, shipped broadly, with dual-use cyber work blocked. Claude Security is the next one. Users receive a defined artifact. They do not prompt Mythos 5.

Enterprise admins turn the feature on in the admin console. From claude.ai/security, a user selects a repository they own. Claude Security scans the codebase and returns each finding tagged for human triage: Common Weakness Enumeration category, confidence, severity, and a suggested patch. Interactive remediation happens in Claude Code on the web. That step uses whatever models the organization already has access to in Claude Code. The Mythos scan itself does not extend Mythos access anywhere else.

Every patch still needs a person. Anthropic is explicit that suggested fixes are not auto-applied. The scan is a discovery and drafting surface. Claude Code on the web is the human-reviewed patch surface.

For security leads already running SAST in CI, the practical difference is scope and interface. You do not get a Mythos chat. You get a scan result. That is also why the billing story is dull on purpose: Mythos 5 inside Claude Security is not a new SKU. It meters like the rest of the Enterprise plan.

Partners, credits, and a wider verification path

The same announcement reaches past Claude Security.

Anthropic is working with cybersecurity technology and services partners to put Mythos 5 inside products defenders already use for security operations, incident response, threat intelligence, detection engineering, and remediation. The intended pattern matches Claude Security: the end user works through a purpose-built interface, Mythos runs in the background on a defined task, and the product returns a specific artifact — a patch list, an alert — rather than a steerable model. Anthropic says it and its partners also have abuse-prevention measures to keep the model inside that scope. The work is early. Security vendors can register interest.

Alongside the product work, Anthropic launched the Defender Advantage Fund, or 0xDAF: $35 million in Claude credits for organizations helping open-source maintainers secure their software. Grants will focus on three areas — patching live vulnerabilities in widely used projects, automating scanning and patching in ways other projects can copy, and backing more ambitious approaches that resist whole classes of attack. The company is starting with a small number of larger pilot grants and says it will name initial recipients in the coming weeks. Project Glasswing had already put $4 million in direct donations plus credits and scanning support behind open-source security, including coordinated efforts such as Akrites and Gold Eagle. 0xDAF is the scaled credit layer on top of that.

The Cyber Verification Program is also expanding. Today it gives vetted defenders reduced safeguards on Opus and Sonnet for authorized work. Over the coming weeks Anthropic says it will widen dual-use capabilities on those models, with Mythos-class access to follow. Glasswing continues in parallel with U.S. government partners for operators of critical infrastructure that meet strict control requirements. Accepted teams do not need to re-apply; everyone else doing legitimate defensive work is invited to apply for the current Opus and Sonnet path.

What this is — and is not

This is not general Mythos 5 chat. It is not an auto-merge bot. It is a scan-only backend inside a public-beta Enterprise product, plus a partner integration track, a credit fund, and a verification-program expansion still weeks away.

For Enterprise teams, the useful test is operational. Enable Claude Security, pick a repo you own, and treat the CWE-tagged findings as a backlog to review — then open Claude Code on the web only for patches a human is ready to own. Watch token usage the same way you already watch the rest of the plan. If you build security products, the partner form is the on-ramp, not a request for raw model access.

Read Anthropic’s August 21 announcement for the official product claims, then decide whether a first scan belongs on a high-risk service or a well-understood repo you can score against your existing SAST baseline.

Findings, not a chat window

Partners, credits, and a wider verification path

What this is — and is not

Sources