> ## Documentation Index
> Fetch the complete documentation index at: https://flox.dev/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Audit an agent plugin

> Scan an agent plugin for security issues with SkillSpector before or after you install it.

A plugin's skills are instructions your agent follows.
Before you install one, you can scan it for prompt injection, credential access, and other risky patterns.
`flox agent audit` runs NVIDIA [SkillSpector](https://github.com/nvidia/skillspector) over each skill and the rest of what the plugin carries, and scores it from 0 to 100, where higher is better.

## Install the scanner

The audit needs the `skillspector` program and network access to `api.osv.dev`, which it uses for live vulnerability data.
Install SkillSpector into your environment:

```bash theme={null}
flox install agent-stacks/skillspector
```

## Audit a package before installing it

Name the catalog package.
Flox fetches it into a temporary environment, audits it, and deletes the temporary environment afterwards:

```console theme={null}
$ flox agent audit agent-stacks/agent-plugin-frontend-design
agent-stacks/agent-plugin-frontend-design → /nix/store/44fwh49l…-agent-plugin-frontend-design-0.0.0-unstable.2026-10-19.ga0b04d9
skillspector 2.12.0 · static · osv live 2026-10-21T15:34Z

Checking
  skills/frontend-design (1 skill)
Validated only
  plugin.json

  skills/frontend-design   87  safe  1 finding

  HIGH    AR2  Anti-Refusal Statement  skills/frontend-design/SKILL.md:69

score 87 · safe (worst: skills/frontend-design)
1 finding: 1 high
```

Add `@<version>` to audit a specific version, for example `agent-stacks/agent-plugin-caveman@4.0.1`.

## Audit what is already installed

Pass a file path instead of a package name.
To audit one installed plugin:

```bash theme={null}
flox agent audit "$FLOX_ENV/share/agent-plugins/frontend-design"
```

To audit every plugin in the active environment and get one overall score:

```bash theme={null}
flox agent audit "$FLOX_ENV"
```

A bare word is always read as a package, so `flox agent audit frontend-design` looks in the catalog, not in your environment.

## Read the result

Each skill and component gets its own score, and the plugin scores its worst.
The bands are SkillSpector's:

| Score | Status |
| - | - |
| 80–100 | `safe` |
| 50–79 | `caution` |
| 0–49 | `do not install` |

Findings are listed with their severity, rule, and the file and line that matched.
HIGH and CRITICAL findings move the score most.
For a `caution` score, read each finding and decide whether the skill's text justifies it.

The audit also says what it did **not** check.
It scans the plugin's skills, its hooks, agents and commands, a `bin/` with programs of its own, and other folders such as `scripts/` and `src/`.
MCP server configuration and files outside the plugin layout, such as `README.md`, are listed under "Not checked", and the score says nothing about them.
Nothing is executed, and no language model reads the plugin.

## Gate on a score in CI

`--threshold` makes the audit exit with status 1 when the score is below the number you give:

```bash theme={null}
flox agent audit "$FLOX_ENV" --threshold 80
```

A threshold of 80 fails on anything that is not `safe`.
On a slow runner, give each scan more time with `--scan-timeout`, for example `--scan-timeout 15m`.

## Related

* [`flox agent audit`](https://flox.dev/docs/ai-agents/reference/flox-agent-audit) reference, including the JSON output and exit codes
* [`flox agent check-plugin`](https://flox.dev/docs/ai-agents/reference/flox-agent-check-plugin), which checks that a plugin follows the specification rather than whether it is safe


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.