Skip to content

SECURITY.md

Latest commit

 

History

History
59 lines (39 loc) · 3.12 KB

SECURITY.md

File metadata and controls

59 lines (39 loc) · 3.12 KB

Security Policy

Reporting a Security Issue

Do not open a public GitHub issue for security vulnerabilities.

If you find a skill that contains malicious content, prompt injection, or any security concern:

  1. Open a GitHub Security Advisory (private by default)
  2. Or email: skills@agentskillexchange.com with subject line: [SECURITY] <brief description>

We will acknowledge receipt within 24 hours and aim to resolve confirmed issues within 72 hours.

What Counts as a Security Issue

Agent skills are instructions that AI agents follow. A malicious skill could instruct an agent to:

  • Exfiltrate data — send files, environment variables, or credentials to external endpoints
  • Prompt injection — override the user's instructions or the agent's safety guidelines
  • Credential harvesting — trick users into providing API keys, passwords, or tokens
  • Malicious URLs — link to phishing sites, malware downloads, or tracking endpoints
  • Destructive commands — delete files, drop databases, or modify system configurations without user consent
  • Privilege escalation — instruct the agent to bypass permission checks or access restricted resources

If a skill does any of these things — whether intentionally or accidentally — report it.

Trust Model

Tier Description
Listed Published — backed by a real tool, repo, or package. Basic format validation.
Security Reviewed Content scanned for prompt injection, data exfiltration, credential harvesting, and malicious patterns. Cleared for production use.

Skills that fail security review are either fixed or removed from the marketplace and this repo.

Security Review Criteria

For a skill to reach Security Reviewed status, it must pass:

  1. No external data transmission — the skill must not instruct the agent to send data to endpoints the user hasn't explicitly configured
  2. No credential requests — the skill must not ask for or handle credentials beyond what its stated APIs require
  3. No instruction overrides — the skill must not contain prompts that attempt to override the agent's system instructions or safety guidelines
  4. No destructive defaults — any commands that modify or delete data must require explicit user confirmation
  5. URL safety — all URLs in the skill must resolve to legitimate, non-malicious destinations
  6. Transparent behavior — the skill must do what its description says and nothing more

Responsible Disclosure

  • We credit reporters in the security advisory (unless you prefer anonymity)
  • We do not pursue legal action against good-faith security researchers
  • We ask that you give us reasonable time to address the issue before public disclosure

Scope

This policy covers:

For security issues with the website itself (XSS, CSRF, authentication bypasses), email skills@agentskillexchange.com with [WEB SECURITY] in the subject.