Skip to content

Security: AgentsKit-io/agentskit

Security

SECURITY.md

Security Policy

Supported Versions

AgentsKit packages are versioned and released independently. Security fixes are provided for the latest version published to npm of each @agentskit/* package. Older versions stop receiving security fixes when a newer version of that package is published.

Package maturity and compatibility guarantees are documented in docs/STABILITY.md. Stable packages follow the documented deprecation cycle for API changes; that cycle does not extend security support to superseded versions.

Reporting a Vulnerability

Please report security issues through GitHub's private vulnerability reporting:

https://github.com/AgentsKit-io/agentskit/security/advisories/new

If that channel is unavailable, email emersonfbraun@gmail.com with [agentskit-security] in the subject.

Please include:

  • Affected package(s) + version(s)
  • Reproduction steps or a minimal proof of concept
  • Suspected severity (see SLA below)
  • Whether the issue has been disclosed elsewhere

Please do not file a public issue or PR for security reports.

Disclosure SLA

Severity First response Patch target Public advisory
Critical (RCE, sandbox escape, credential exfiltration) 24 h 7 days within 7 days of patch
High (SSRF, auth bypass, data exfiltration) 48 h 14 days within 14 days of patch
Medium (DoS, info leak, hardening gap) 5 days next minor release with release notes
Low (best-practice, defense-in-depth) best effort next minor release with release notes

The patch target is "design + implement + test + publish to npm". For ecosystem-wide issues we coordinate with downstream maintainers before public advisory.

Scope

Covered by this policy:

  • All @agentskit/* npm packages published to the npm registry
  • The agentskit CLI binary
  • This repository's GitHub Actions workflows

Out of scope (best-effort only):

  • Example apps under apps/example-*
  • Third-party adapters published outside the @agentskit scope
  • The docs site (apps/docs-next) hosted on Vercel

Dual-use components

The following packages and modules ship intentionally powerful primitives. They are designed for sandboxed / supervised use; running them against untrusted input without the documented hardening is itself a misconfiguration, not a vulnerability.

  • @agentskit/toolsshell, filesystem, fetchUrl, sqliteQueryTool, mcp client
  • @agentskit/sandbox — code execution backends
  • @agentskit/core/security — PII vault, injection detector, rate limiter

Each module's README documents the recommended deployment posture (allowlists, sandbox wrapping, audit sinks). Vulnerability reports about these modules should describe how the documented guarantees were broken, not that the underlying capability exists.

Supply chain

  • All third-party GitHub Actions are pinned to commit SHAs.
  • actions/dependency-review-action runs on every PR; high-severity vulnerabilities or copyleft-licensed deps fail the build.
  • pnpm audit --prod --audit-level=high runs in CI.
  • CodeQL (security-and-quality query pack) runs on every PR plus a weekly schedule.
  • SBOMs are generated on release; ask if you need a specific format.

Coordinated disclosure

We follow standard 90-day coordinated disclosure unless severity or exposure justifies a shorter window. Reporters are credited in the advisory unless they request otherwise.

There aren't any published security advisories