---
source: crewkit documentation
url: https://crewkit.io/docs/skill-scanning.md
title: "Skill and agent security scanning"
description: "What crewkit checks before publish, what a grade means, and how adoption gates first-party catalog items."
---


Before a skill or agent is published, and again when its content changes, crewkit scans the text that would reach a developer's `.claude/` tree. A letter grade summarizes the findings. That grade is one input to publishing and serving — not a guarantee that a resource is safe or high quality.

Pattern counts below were read from the scanner at ship time (**November 20, 2026**). The catalog grows as detectors are added, so read these counts as a snapshot, not a fixed total.

---

## Static scan

The static scan is a **pattern check** over the resource body plus its name, description, and frontmatter — the same material crewkit would write into the repository. For a playbook, convention text is included too. It looks for risky shapes in six categories:

| Category | What it targets | Patterns (at writing) |
|----------|-----------------|----------------------|
| **secrets** | Hardcoded credentials and token-like strings | 9 |
| **injection** | Shell execution and command substitution | 7 |
| **unsafe_url** | Non-HTTPS or unusual URL forms | 3 |
| **prompt_injection** | Instructions that try to override or bypass agent behavior | 16 |
| **path_traversal** | Paths into sensitive system or home directories | 5 |
| **exfiltration** | Network calls that may send data outward | 4 |

**Total: 44 patterns in 6 categories.**

For the publish and serve grade, the static pass runs on the text as written and again on a sanitized form that strips invisible and bidirectional characters; the worse of the two forms sets the static portion of the grade.

When someone validates a skill in the Builder, the same pattern catalog runs over the draft and supporting files. Any **critical** finding blocks validation.

---

## Model analysis

A second pass sends a truncated copy of the resource — name, description, frontmatter, body, and for playbooks the convention text, up to 16,000 characters — to **Claude Haiku 5.5** for semantic review. It is meant to catch issues pattern matching may miss, such as hidden instructions, unsafe operations, and privilege-escalation wording.

**By default**, crewkit runs this pass for your organization's own skills and agents. It **does not** send that content out when the organization has any of these on:

- [Data residency](/docs/data-residency) (a resident organization)
- [Capture mode](/docs/capture-modes) set to **minimal** (privacy mode)
- Content egress turned off (see [Capture modes](/docs/capture-modes) and the organization rows on the [settings map](/docs/settings-map#organization))

When the model pass does not run, the grade reflects static findings only and the stored verdict records that fact.

**crewkit catalog** (platform) skills and agents get the model pass when the service is configured and the static scan finishes without timing out. Publishing a catalog item also requires a completed model verdict. If the static scan times out twice, the model pass is skipped and the item is graded **F**.

If the deployment has no Anthropic API key configured, the model pass does not run and the stored verdict says so.

---

## Letter grades and what passes

Grades are **A** through **F**. crewkit assigns a grade from the combined static and model findings:

| Grade | Rule (severity counts) |
|-------|-------------------------|
| **F** | Any critical finding |
| **D** | Two or more high findings |
| **C** | One high finding, or three or more medium findings |
| **B** | One or two medium findings |
| **A** | No critical, high or medium finding (low findings alone never lower the grade) |

**A**, **B**, and **C** are passing. The **Publish** action on the dashboard requires a passing grade. Managers can still sync **draft** resources that were never scanned when they use draft sync; those drafts are not served to every teammate until published and graded.

Serving excludes only resources **known** to be below the passing bar (stored **D** or **F**).

---

## When scans run

| Trigger | What runs |
|---------|-----------|
| **Publish** | Full pipeline when the content changed or was never analyzed: static (raw + sanitized), then the model pass when permitted. A grade below **C** blocks publish. |
| **Published content edit** | The same full pipeline runs again in the background. A failing grade can unpublish the resource. |
| **Serve / CLI sync** | Published resources with a stored grade of **D** or **F** are withheld from sync. Resources with no grade yet still sync until a scan completes. |
| **Skill Builder validate** | Static scan only; critical findings return validation errors. |

---

## What a scan keeps

For each finding, a completed scan stores the category, severity, line number when available, and a description on the resource until the next scan replaces them. A pattern finding's description is crewkit's fixed wording for that pattern — prefixed with the convention label and recorded at line 0 when it matched a playbook convention — and a static scan timeout records a fixed timeout message with no line number. A model finding's description is the model's own wording, with recognised secret shapes replaced by a redaction marker and cut to 500 characters before it is stored. Descriptions stored by earlier scans are not rewritten: they stay on the resource until its next completed scan, and in audit rows as described below. Unlike that model-written description, the text a pattern matched is stripped before saving and is not stored on the resource.

The current grade and finding list stay on the resource itself until a later scan overwrites them. Content edits usually create an audit row as well. A completed scan's result, its findings and their descriptions included, is usually recorded in an audit row too, so a description that a later scan replaced on the resource stays in the audit rows for as long as the next sentence says. For resources, crewkit's daily retention job removes audit rows once they are older than **one year**; for a resource that has not been deleted, it keeps the newest audit row and, once that row is older than one year, clears its copy of the earlier content and its record of what changed. While a backlog drains, this can take several daily runs. Some automated scan outcomes are written without a new audit row.

---

## What a grade does and does not mean

A passing grade means crewkit's automated checks did not find enough severity to block publish or to treat the resource as known-insecure on sync. It does **not** mean:

- The resource is free of malicious intent or subtle runtime risk.
- The model pass ran (egress may be barred, the analyzer may be unconfigured, the static scan may have timed out, or a catalog item may be waiting on a failed retry).
- Every line was reviewed by a human.
- The grade already reflects an edit you just saved — until the background rescan finishes, sync may still use the previous grade.

Treat the grade as a gate crewkit applies consistently, not as certification.

---

## Malware scanning

Antivirus scanning applies to **uploaded document bytes** and **ask images**, not to skill or agent resource text. It is unrelated to the letter grade on a skill or agent.

---

## Adoption gate (first-party catalog)

Separately from the grade, a **crewkit catalog** (first-party platform) agent or skill reaches your projects only after your organization has opted in.

**Organization scope —** in the sidebar, open **Skills**, then **Catalog** (`/kit/skills`). Turning a skill or agent **On** here adopts it for the whole organization (every project unless a project overrides it).

**Project scope — the project's Skills page**. A Manager or above can turn a catalog skill **On** for that project only, **Off** for that project, or **Use organization default**.

Until you adopt, catalog seeding does not write those resources into your repositories. Adoption does not replace scanning: catalog items are still graded before publish and on content changes. Adoption controls **who opted in** to receive a first-party item, on top of the security verdict.

New organizations start with the **crewkit-guide** agent already adopted organization-wide, plus the default session workflow skills from crewkit's onboarding setup.

---

## Next steps

- [Skills](/docs/skills) — turn catalog skills on for your team
- [Agents](/docs/agents) — create and sync agents
- [Inheritance](/docs/inheritance) — platform, organization, and project tiers