More screenshots
Works with the AI Agents you already use
Agent Skill Supply Chain Audit
Audit an AI agent skill package before you install it. Finds unsafe shell commands, embedded secrets, obfuscated payloads, prompt-injection text and undeclared permissions, with a 34-rule catalogue and JSON evidence.
Free
Agent Skill Supply Chain Audit
Example session with this skill installed
I downloaded a skill package from a marketplace and I want to know whether it is safe to install. Audit the candidate package at ./vendor-weather-skill and tell me what is wrong with it before I run anything.
- Read your context and instructions
- Compiled the agent supply chain
Audit Findings Report: vendor-weather-skill
Target: vendor-weather-skill
Files scanned: 4 | Total size: 540 bytes | SKILL.md files: 1
Findings: 5 | Highest severity: critical
{
"tool": "skill-package-auditor",
"version": "1.1.0",
"target": "vendor-weather-skill",
"summary": {
"fileCount": 4,
"totalBytes": 540,
"skillMdCount": 1,
"maxSeverity": "critical",
"findingCount": 5
},
"manifest": {
"path": "SKILL.md",
"hasFrontmatter": true,
"name": "Weather Helper",
"description": "Fetches the current weather for a city and caches it locally.",
"version": "",
"allowedTools": [
"*"
],
"declaredPermissions": [],
"observedCapabilities": {
"terminal": false,
"read-files": false,
"write-files": false,
"environment": false
}
},
"observedCapabilities": {
"terminal": false,
"read-files": false,
"write-files": false,
"environment": false
},
"findings": [
{
"id": "SHELL-004",
"category": "shell",
"severity": "critical",
"path": "scripts/setup.sh",
"line": 3,
"title": "Force delete of an absolute or home path",
"evidence": "rm -rf /var/tmp/weather-cache",
"remediation": "See SHELL-004 in references/rule-catalogue.md"
},
{
"id": "DEP-001",
"category": "dependencies",
"severity": "high",
"path": "requirements.txt",
"line": 0,
"title": "Dependency version not pinned exactly",
"evidence": "requests>=2.0",
"remediation": "See DEP-001 in references/rule-catalogue.md"
},
{
"id": "MAN-001",
"category": "manifest",
"severity": "high",
"path": "SKILL.md",
"line": 1,
"title": "Wildcard tool grant in frontmatter",
"evidence": "allowed-tools: *",
"remediation": "See MAN-001 in references/rule-catalogue.md"
},
{
"id": "SEC-006",
"category": "secrets",
"severity": "high",
"path": "config/env.sh",
"line": 1,
"title": "Generic prefixed API key",
"evidence": "credential-shaped value on this line",
"remediation": "See SEC-006 in references/rule-catalogue.md"
},
{
"id": "MAN-002",
"category": "manifest",
"severity": "low",
"path": "SKILL.md",
"line": 1,
"title": "Frontmatter missing a version",
"evidence": "version field absent",
"remediation": "See MAN-002 in references/rule-catalogue.md"
}
],
"fileInventory": [
{
"path": "SKILL.md",
"bytes": 281,
"class": "document"
},
{
"path": "requirements.txt",
"bytes": 22,
"class": "manifest"
},
{
"path": "config/env.sh",
"bytes": 117,
"class": "script"
},
{
"path": "scripts/setup.sh",
"bytes": 120,
"class": "script"
}
]
}
Recommendation
Do not install this package as it stands. The setup script force-deletes an absolute
path, the manifest grants every tool while the code itself declares none, a config file
carries a credential-shaped value, and the one dependency is not pinned. Fix the shell
rule, narrow the tool grant, remove the credential and pin the dependency, then re-run
the audit before you run anything.
Next steps
- Open
scripts/setup.shand remove the force-delete command on line 3. - Replace the wildcard tool grant in the SKILL.md frontmatter with the exact tools the code uses.
- Remove the credential from
config/env.shand read it from the environment instead. - Pin
requeststo an exact version inrequirements.txt. - Add a
versionfield to the SKILL.md frontmatter. - Re-run the audit and confirm it reports zero findings before you install the package.
Connects securely to your tools. The creator never sees your data.
About this skill
An agent skill is a folder of instructions and code that runs with your system privileges. When you download a skill from a marketplace or a repo you do not control, you are inviting a third party to execute commands and read files on your machine. Standard code reviews often miss the specific risks of instruction-override attacks and over-broad manifest permissions.
What it does
- Manifest validation compares declared
allowed-toolsagainst actual file behavior to find permission gaps. - Static analysis runs a deterministic, offline scan for unsafe shell commands, hardcoded secrets, and obfuscated payloads.
- Injection detection identifies attempts to redirect or weaken the agent instructions within the
SKILL.mdbody. - Dependency auditing flags unpinned or untrusted dependencies that could compromise the supply chain.
- Ranked reporting generates a verdict-based report with evidence, severity levels, and specific remediation steps.
How it works
- Scope definition captures the package directory and content digest to ensure you are auditing the correct artifact.
- Manifest extraction reads the frontmatter to establish a baseline of what the skill claims to do versus what it asks for.
- Structural analysis runs the bundled Python analyzer to generate a JSON inventory of permissions and file types.
- Rule matching applies 34 specific rules across shell, injection, and secret categories to find matches.
- Verdict generation produces a final report (Install, Install with changes, or Reject) with redacted evidence.
Frameworks & tools
This skill uses a bundled Python structural analyzer. It relies on standard filesystem tools like read, grep, and glob to perform offline, deterministic audits.
Why this beats prompting it yourself
Generic prompts often treat instructions as documentation rather than executable code. This skill enforces a 34-rule security catalog and a specific severity model that ensures consistent, citeable findings instead of vague "it looks okay" summaries.
Use cases
- Vetting third-party marketplace skills before local installation.
- Auditing skill changes in a CI/CD pipeline or pull request.
- Screening untrusted downloads for embedded credentials or obfuscated code.
- Building a standardized security review checklist for teams publishing internal skills.
Known limitations
Pattern matching only catches known attack shapes, not novel exploits. It does not model runtime agent behavior or recognize every possible obfuscation encoding.
Detects known patterns, not novel attacks. Does not model runtime agent logic. Probes terminal, file, and env access but requires manual review for network calls.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 4 days ago
- Free with an Agensi account
Requires
Permissions
File Scopes
The analyser reads the target package as text and writes nothing. It never imports, executes or installs anything it finds.