Anthropic spent months gating Claude Mythos 5 behind export controls and Glasswing approvals. Security teams wanted the capability. Compliance teams wanted guardrails.
The August 21, 2026 answer is a product shape, not a model dump: Claude Security now runs Mythos 5 scans for every Enterprise customer, but you never get a chat window into the cyber model itself.
What shipped
From Anthropic's defender update:
| Piece | Detail |
|---|---|
| Scanner | Claude Security on claude.ai/security |
| Model | Mythos 5 inside the scan harness |
| Input | GitHub repos via Anthropic GitHub App |
| Output | CWE category, confidence, severity, suggested patch |
| Billing | Standard Enterprise token usage, no separate SKU |
| Bonus | $35M Defender Advantage Fund credits for open-source security work |
The security story is outputs, not access. Anthropic argues defenders can use frontier cyber capability when the user only receives patches and alerts, not an steerable Mythos prompt surface.
How this differs from the Claude Code security plugin
I covered the terminal stack in Claude Code's security plugin. Keep the lanes separate:
| Surface | Who | Model | Where it runs |
|---|---|---|---|
| Claude Security (hosted) | Enterprise admins | Mythos 5 in scan harness | Anthropic cloud + your GitHub |
| Claude Security plugin | Claude Code users (beta) | Models on your plan | Your laptop session |
| Security guidance hook | Claude Code users | Pattern matcher (no model) | Local, zero token cost |
If you are on Pro or Max, the plugin is your path today. If you are Enterprise with GitHub org repos, Mythos scans are the new tier.
Setup checklist (Enterprise)
Anthropic's docs and partner writeups converge on this gate:
- Enterprise contract with Claude Security enabled by an org admin
- Claude Code on the web + Extra Usage turned on where required
- Anthropic GitHub App installed with repo scope
- Premium seat for engineers who launch scans
- Navigate to claude.ai/security, pick a repo, run the scan
Each finding should include:
- CWE label (standard weakness taxonomy)
- Confidence + severity for triage ordering
- Suggested patch you open in Claude Code on the web for human review
Patches are never auto-applied. That is intentional.
Why Mythos behind glass matters
Mythos-class models sit in a awkward spot: high defensive value, high misuse risk if prompt-steered into exploit generation.
Anthropic's pattern here mirrors what I wrote about export controls on Fable and Mythos: capability through products, not raw weight downloads.
For CISOs, the question is not "can my devs access Mythos?" It is "can Mythos inspect our repos without becoming a general-purpose attack assistant?" Claude Security's answer is scoped scan artifacts.
What I would pilot first
Do not wire this into CI on day one. Public beta means false positives and cost surprises still happen.
Week 1 pilot:
- Two repos: one mature service, one greenfield agent-heavy branch
- Compare Mythos findings against your existing SAST + SkillSpector skill scans
- Track CWE distribution and human acceptance rate on suggested patches
Week 2:
- Add a pre-release scan for repos with heavy agent-generated diffs (see agentic coding bottlenecks)
- Document which finding classes actually block merges vs create noise
Cost and governance
Scans consume Enterprise tokens like any other Claude workload. That is cheaper than a standalone AppSec SKU until volume spikes.
Governance hooks I recommend:
| Control | Why |
|---|---|
| Repo allowlist | Stop "scan everything" token burns |
| Human merge gate | No auto-apply patches from beta scanner |
| Finding export to SIEM | CWE-tagged JSON plays nice with existing triage |
| Separate from prod credentials | Scanning service accounts should be read-only on code |
Open-source angle: the $35M fund
The Defender Advantage Fund (0xDAF) puts $35M in Claude credits toward patching vulnerabilities in open-source software. If you maintain OSS dependencies your clients rely on, watch Anthropic's application path. It pairs with the Cyber Verification Program expansion announced in the same post.
Bottom line
Mythos 5 moved from "approved defenders only" to "Enterprise scan button" without opening a public cyber chat API. That is the right packaging for most companies.
If you need help stitching hosted scans, local Claude Code plugins, and agent CI policy into one playbook, book a free discovery call. Security for agentic shipping is half tooling, half process.