SatpamAI (satpam is Indonesian for “security guard”) is a Claude Code skill that acts as a mandatory checkpoint before the AI runs any other skill. The premise: third-party AI skills are untrusted artifacts that can hide prompt injection, malware, hardcoded secrets, or supply-chain compromise — so inspect first, act last. Its motto: “No artifact earns trust. Trust is proved by evidence.”
It runs an 8-phase security gate (classify → threat-assess → inspect → decide → controlled execute → verify → audit) and a 10-layer malware analysis (provenance, static code analysis, secret scanning, dependency and permission auditing, integrity checks). A dedicated engine hunts for prompt injection — override attempts, authority impersonation, bypass requests, obfuscated multi-vector payloads — and everything gets a risk rating from PASS up to CRITICAL / UNKNOWN. Crucially it is static-only (“never execute untrusted code to test if it is malicious”) and it fails closed: uncertainty blocks.
One design choice worth noting: “prior session approval means nothing here” — each decision stands on its own evidence, with no grandfathered trust. Install via claude skill install <repo> (or /satpamai, or auto-load from CLAUDE.md). From the same author as SANITY; MIT licensed.
source: github.com/verycoolapps/SatpamAI
my other side: ivo.blogspot.com
Leave a Reply