Claude Enterprise adds beta security scanning for third-party skills and plugins

If your organisation uses third-party skills or plugins with Claude, there is now an automated way to check them for malicious content before they go anywhere near your users. Anthropic has launched beta skill and plugin security scanning for Claude Enterprise plans, giving admins an opt-in control that automatically reviews every third-party skill or plugin at the moment it is uploaded or edited.

What is actually being scanned, and why does it matter?

Skills and plugins extend what Claude can do. They can connect to external services, access files, or execute code on behalf of your users. That is exactly what makes them useful, and exactly what makes a malicious one dangerous.

Anthropic’s own documentation is candid about the risk: “Never deploy Skills from untrusted sources without a full audit.” A compromised skill could instruct Claude to execute arbitrary code, access sensitive files, or quietly send data to an outside server. The two attack patterns Anthropic flags most prominently are prompt injection, where a skill manipulates Claude into taking unintended actions, and data exfiltration through malicious package code or injected instructions.

In practice, manually auditing every skill upload in a large organisation is not realistic. That is the gap this feature addresses.

How the scanning works

When an admin enables skill and plugin security scanning, Claude automatically inspects each third-party skill or plugin as it is uploaded or edited. The scan looks for signs of malicious behaviour: hidden code execution, instructions that route your data to an outside service, or content that attempts to tamper with Claude’s safeguards.

Each scan returns one of three results:

Pass. Nothing concerning was found. The skill or plugin installs and works normally.

Warn. The scan could not fully verify the skill, and it may carry risk. The skill remains usable, but a caution banner appears that the user must acknowledge before continuing.

Fail. The scan detected malicious content. The skill or plugin is blocked immediately and cannot be used. The banner explains the specific reason it was flagged. A blocked skill cannot be overridden by the uploader and cannot be approved for the organisation at this point. The only path forward is to fix the flagged issue and re-upload.

Most scans complete in roughly one to two minutes and run in the background, so they do not create significant friction for legitimate uploads.

Turning it on

The feature is off by default. Owners and Primary Owners can enable it by going to Organization settings → Skills in claude.ai and toggling on “Skill and plugin security scanning.”

Once enabled, scanning applies across the whole organisation and covers new uploads and edits in claude.ai, Claude Cowork, and Enterprise plugin marketplaces. Skills and plugins already installed before you enable the feature keep working as they were; the scan only runs on new activity from that point forward.

If your organisation uses custom roles, there is an additional layer of control. Admins can enable the skill and plugin security scanning capability for specific roles, allowing more granular decisions about where scanning applies rather than a blanket organisation-wide setting.

There is no extra cost. Scanning is included in the Enterprise plan.

What this means for you

For IT admins and security teams, this removes a manual step that was easy to skip and hard to enforce consistently. Rather than relying on individuals to vet the skills they install, the platform checks them automatically. A failed scan blocks the skill before anyone can use it, without requiring an admin to intervene in real time.

For end users, the experience is mostly invisible. A skill that passes installs as normal. A warn result means you will see a caution message to read before proceeding. A fail means the skill is unavailable, and the person who uploaded it will need to resolve the issue.

For security-conscious organisations that have been cautious about enabling skills at all, this lowers the barrier meaningfully. You still need governance policies around which skills are approved and who can upload them, but you now have an automated backstop that checks each one for known malicious patterns before it runs.

Part of a broader security push from Anthropic

This feature sits alongside other security investment Anthropic has been making at the platform level. On 21 July 2026, Anthropic brought Claude Security, its multi-agent vulnerability scanner, directly into Claude Code as an installable plugin in public beta. That targets code-level vulnerabilities; skill and plugin scanning targets the supply chain risk of third-party extensions. The two sit at different layers of the stack but reflect the same direction of travel.

Anthropic’s product framing for Claude Security describes a three-step workflow: scan, validate, patch. Skill and plugin scanning fits the scan step for the extensions layer of Claude Enterprise, adding automated inspection where organisations were previously reliant on manual review or trust.

How to get started

If you are on a Claude Enterprise plan, head to Organization settings → Skills and enable the scanning toggle. From that point, every new skill or plugin upload or edit in your organisation will be checked automatically.

The feature is in beta, so Anthropic will likely iterate on the scanning logic and result categories over time. Given that it is free and opt-in, there is little reason to wait if you are managing skills at any scale.