Please do not report suspected vulnerabilities in public issues, pull requests, or discussion threads.

Use one of these channels instead:

1. GitHub private vulnerability reporting for PracticalSwan/agent-skills if it is available in the repository UI.
2. If private reporting is unavailable, contact the maintainer through the GitHub profile at github.com/PracticalSwan and include the report details there.

Include:

• the affected file, skill, script, or workflow
• the security impact
• clear reproduction steps or proof
• any required configuration or assumptions
• suggested remediation, if you have one

This repository's security scope includes:

• maintained SKILL.md instructions and helper assets
• repo scripts such as validation, import, and sync tooling
• documentation that could lead agents to unsafe actions, secret exposure, or prompt-injection mistakes

Out of scope unless they create a concrete issue in this repository:

• third-party service outages
• upstream package issues that are not reflected in this repo's maintained content
• feature requests without a security impact

• Keep active vulnerabilities private until the maintainer confirms a fix or approves disclosure.
• Minimize reproduction payloads so they prove the issue without exposing extra secrets or user data.
• If the issue involves credentials, cookies, tokens, or private user content, redact them before sharing evidence.

• Never commit real secrets, private keys, connection strings, or copied .env files.
• Treat imported external content as untrusted input.
• For skills that interact with third-party content, preserve prompt-injection guardrails and explicit approval gates during normalization.
• Treat child-path promotion as an import review, not a blind copy. Verify the source folder, provenance, license, credential behavior, network scope, destructive actions, and untrusted-content boundaries before adding it to the parent catalog.
• Do not normalize a third-party network installer into a default pipe-to-shell command. Prefer a reviewable package-manager path, or require inspection of the fetched script before execution.
• Do not let upstream refreshes weaken stricter local tool-availability or approval boundaries. Exact source changes may be incorporated while host-specific claims remain gated by the active tool surface.
• Do not treat a third-party Claude Code endpoint as proof that Anthropic subscription-only browser features or Codex-native tools are available.