Claude Security Scans Now Run on Claude Mythos 5

Anthropic said on 21 August 2026 that Claude Security scans now run on Claude Mythos 5. The product is in public beta for Claude Enterprise. There is no separate add-on. Scans bill as standard token usage.

How Claude Security uses Mythos 5 without a prompt box

An admin turns the feature on in the console. A user then goes to claude.ai/security, connects a GitHub repository, and runs a scan. Mythos 5 is not handed over as a prompt box. The scan returns findings. It does not open a general chat with the cyber-capable model.

That scoping is the product. Teams get Mythos-quality vulnerability results without a new interactive surface. If your org does not already have Mythos access, Claude Security still does not give you a place to type at it.

The Claude Security plugin for Claude Code is in beta. That plugin uses only the models already in that account. It does not use Mythos. Interactive patching stays on whatever you already pay for in Claude Code.

Connecting a GitHub repo also means the scan depends on GitHub staying up. A GitHub outage is a scan outage, not a Claude outage. Keep that distinction in the runbook.

Claude itself had a degraded stretch on 18 August and an outage on 20 August that hit the app, API, Code, and Cowork. Claude Security is another Anthropic surface. Do not treat it as an independent spare scanner.

What a scan returns, and who can apply a patch

Each finding comes back with a CWE category, a confidence rating, a severity rating, and a suggested patch. Anthropic also runs an adversarial verification pass: the system challenges its own results before they surface.

A human must approve patches. Nothing lands because the model liked its own fix. That is the control that matters more than the model name. If your process auto-merges agent diffs, Claude Security is not a license to keep doing that.

Suggested patches are still model output. Review them the way you review any generated change, especially on auth, crypto, and anything that touches untrusted input. The product is a scanner plus a proposed diff, not a signed-off security exception.

Put the workflow in a skill prompt: who may start a scan, which repos are in scope, who approves a patch, and which model actually applies it in Claude Code. Do not bury "use Mythos" in a paragraph. Mythos is on the scan path only.

Credits and an expanding verification program

Unite.AI reported the same announcement. Anthropic also announced $35 million in credits for an open-source defense fund. The Cyber Verification Program is expanding. Those are adjacent moves, not extra switches inside Claude Security.

Do not invent a price for Mythos scans beyond the stated billing: standard token usage, no separate add-on. Do not invent who receives the credits. The figure is $35 million in credits, attributed here to Unite.AI.

For prompt and skill authors, the operational change is narrower than the headline. Enterprise customers can scan a connected GitHub repo at claude.ai/security, on Mythos 5, after an admin enable. Findings arrive tagged. Patches wait on a person. Claude Code's security plugin, in beta, stays on the models already in the account.

That is enough to update a security skill this week. It is not enough to claim that every Claude surface now runs Mythos 5.

Sources

• Claude Security — Anthropic product page
• Anthropic Deploys Claude Mythos 5 in Security Tools, $35M Open-Source Fund — Unite.AI, covering the 21 August 2026 announcement