Releases: empowered-humanity/agent-security
Releases · empowered-humanity/agent-security
v2.0.0 — Runtime Guards + 30 New Patterns
What's New in v2.0.0
30 New Detection Patterns (190 → 220)
- Infrastructure attack patterns (18 patterns) — env variable injection, symlink traversal, Windows exec evasion, network/fetch misconfig, extended SSRF, bind/proxy misconfig
- Auth anti-patterns (4 patterns) — fail-open catch, string "undefined" coercion, partial identity matching, non-constant-time comparison
- Supply chain install patterns (4 patterns) — curl/wget/PowerShell pipe-to-shell, password-protected archive evasion
- Container misconfiguration (4 patterns) — home dir bind mount, root filesystem mount, seccomp/apparmor unconfined
5 Runtime Guard Modules
Drop-in security middleware for AI agent runtimes:
- SSRF Guard (
guards/ssrf) — DNS pinning, IP blocklists (RFC 1918, loopback, link-local, CGNAT, IPv6), redirect validation - Download Guard (
guards/download) — body size caps, connection/response timeouts, content-type validation - Exec Allowlist (
guards/exec-allow) — default-deny execution, resolved binary path matching, env var filtering, evasion detection - Path Traversal Validator (
guards/fs-safe) — TOCTOU-safe file open, symlink validation, post-open inode verification - Webhook Verifier (
guards/webhook) — timing-safe HMAC verification for GitHub, Slack, Stripe + generic factory
Stats
- 220 detection patterns (was 190)
- 254 tests passing (was 123)
- 5 runtime guard modules (new)
- Full OWASP ASI mapping (ASI01–ASI10)
Research Foundation
All patterns trace to published research:
- OpenClaw vulnerability catalog (80+ security commits, 12 categories)
- ClawHavoc supply chain campaign (341 malicious skills)
- Gemini deep research analysis (45 sources, 8 CVEs)
See CHANGELOG.md for full details.