Use a current Skillware release to receive security fixes. We patch vulnerabilities only for supported versions.

Thresholds are defined in skillware/version_policy.py (MIN_SECURITY_SUPPORTED, MIN_UNSUPPORTED) and bumped by maintainers when support windows change.

Note: PyPI releases are immutable. Users on very old wheels will not see the CLI advisory until they upgrade to a release that includes this logic at least once. That is expected for OSS packaging.

Loading a skill runs its skill.py in your host process. Skillware implements a Permissive Fortress architecture to safeguard the runtime:

1. Credential Sandboxing: Bundled and third-party skills must never read raw os.environ for API keys, secrets, or RPC endpoints. Skills resolve credentials via BaseSkill.credential(key) or SkillContext(secret_provider=...), ensuring host-injected scoping and tenant isolation.
2. Automated AST & Security Scanning: Automated security gates in CI (bandit, tests/test_security_audit.py) prohibit dangerous execution primitives like eval(), exec(), compile(), and dynamic code generation across all bundled skill execution paths.
3. Dependency Auditing: Dependencies are pinned and scanned against known vulnerability databases on every PR and release candidate (pip-audit).
4. Provenance & Review: While skills execute in-process without OS-level sandboxes (like containers or WASM), trust is reinforced by strict human line-by-line review, provenance tracking, and explicit licensing.

Before loading external skills or designing host integrations, review the Skill Trust Model (especially Section 10: The Permissive Fortress Architecture).

We take security seriously. If you discover a vulnerability in Skillware (e.g., standard library skills leaking data, or loader bypasses):

1. Do NOT create a public GitHub issue.
2. Email us at security@arpacorp.net (or contact a maintainer directly).
3. Include a proof of concept if possible.

We will acknowledge your report within 48 hours and provide a timeline for a fix.