Claude Security scans code with new Mythos 5 model

Claude Security is a public beta for Claude Enterprise customers. There's no need to separately apply for access to the Mythos 5 model, but it isn't usable out of the box either. An organization admin must activate Claude Security and Claude Code on the web, turn on Extra Usage, and install the Anthropic GitHub App. Users who actually run scans also need a premium seat within the Enterprise organization. Mythos 5 scans are billed using standard token usage under existing plans, with no separate product fee.

How the scan works

When a connected GitHub repository is selected in Claude Security, Mythos 5 doesn't look at files in isolation — it traces the flow of data across files and infers how components interact with one another. Rather than catching syntax errors in a single function, it follows how a value passed from module A is used in module B to find vulnerabilities.

Each finding comes with the following four pieces of information.

Item	Description
CWE category	Classification code for the common vulnerability type
Confidence	Rating of how likely the detection reflects an actual issue
Severity	Rating of the risk level if the issue occurs
Suggested fix	A proposed patch that can be applied to the code

Selecting a suggested patch opens a fix session in Claude Code on the web. This fix work uses whatever model the organization already has access to in Claude Code. In other words, vulnerability detection is handled by Mythos 5, while the actual code fix is carried out with a Claude model the organization already uses. Patches are not applied automatically — a human must review and approve them before implementation.

The official guidance currently specifies the Claude Security web interface and Claude Code on the web as the access paths. While regular Claude Code can be used in the Claude desktop app, there's no indication that the same Mythos 5 scanning interface is available there.

How to try it

Before getting started, you need an Enterprise account, Claude Code on the web, Extra Usage, the Anthropic GitHub App, and premium seat conditions in place, and an org admin must turn on Claude Security. The process then goes as follows.

1. An admin activates Claude Security and Claude Code on the web, and grants the Anthropic GitHub App access to the target repositories.
2. The user selects a connected GitHub repository at claude.ai/security. The scope can be narrowed to a specific branch or directory if needed.
3. Mythos 5 runs the scan in the background, tracing data flow across files.
4. Once the scan finishes, findings are listed with a CWE category, confidence and severity ratings, and a suggested fix attached to each.
5. Selecting an item opens a fix session in Claude Code on the web, where a human reviews and approves the suggested patch.

It can also be used for scheduled scans to check for data-flow vulnerabilities before new deployments. However, it currently only supports repositories hosted on GitHub, and should only be used on code that you or your company own or have the right to scan. It's not intended for scanning third-party open-source projects without authorization.

Why the model stays hidden behind the scan

Anthropic designed the system so users don't interact with Mythos 5 directly, but instead only receive fixed outputs like vulnerability lists or patches. Opening up direct access could let a malicious user steer the model toward offensive tasks, but running the model behind a purpose-built interface can deliver defensive outcomes while reducing the potential for misuse.

Anthropic also said it's working with partners to integrate Mythos 5 into security products and services. In these cases too, the design keeps users from conversing directly with Mythos 5 — instead, the product runs fixed tasks in the background and delivers only outputs such as a list of patches. A new fund for supporting open-source security, the Defender Advantage Fund, has been allocated $35 million in Claude credits.

The official announcement does not describe this launch as a follow-up to the early-August incident involving unauthorized access to enterprise systems using Claude. What Anthropic has stated directly about this structure is not that it limits the scope of the model's autonomous actions, but that withholding direct prompt access to Mythos 5 from users reduces the potential for misuse of its offensive capabilities.

Editor's take

The notable point in this announcement is less about model performance and more about the access model. Anthropic isn't opening up Mythos 5's offensive cyber capabilities to general users directly, while still delivering defensive outputs — vulnerabilities and patches — to Enterprise customers. It's a choice to broaden the utility of the capability while restricting direct access to it.

For security teams putting a code scanner into practice, deciding which detected issues to fix first is often the hard part. Providing CWE category, confidence, and severity together helps with that prioritization — teams can review patches for items with both high confidence and high severity first, and push lower-priority items back.

For companies in Korea looking to trial this, checking for an Enterprise contract alone isn't enough. An org admin needs to turn on Claude Security and Claude Code on the web, and confirm Extra Usage, GitHub App, and premium seat requirements. After that, it's realistic to run a trial scan on one or two repositories before deployment and check the distribution of CWE categories and cost. Since this is still a public beta, results shouldn't be wired directly into automated deployment pipelines — a human should check confidence, severity, and the suggested patch before applying it.

The next thing to watch is the Mythos 5 integration into partner security products and services that Anthropic has previewed. What defensive outcomes security tools outside Claude Security will deliver without direct model access is the next stage of this trend.