Skip to content

Security Guide

IpMan includes built-in defense against malicious AI agent skills.

Why This Matters

  • 36% of ClawHub skills contain prompt injection (Snyk ToxicSkills)
  • Skills run with full agent permissions, no sandbox by default
  • Attack vectors: credential theft, data exfiltration, obfuscated code execution

Risk Levels

Level Meaning Examples
LOW No issues detected Notes, formatting, weather
MEDIUM Some concerns File operations, API calls
HIGH Red flags detected Credential access, system modification
EXTREME Likely malicious Root access, data exfiltration

Security Modes

ipman install web-scraper --security strict
Mode LOW MEDIUM HIGH EXTREME
permissive Install Install Install Warn
default Install Install Warn Block
cautious Install Warn Block Block
strict Install Confirm Block Block

Set default mode in ~/.ipman/config.yaml:

security:
  mode: cautious

Trust Model

Source Default behavior
IpHub Trust existing risk label
Local .ip.yaml file Mandatory local scan
URL / custom hub Mandatory local scan

Override with flags:

ipman install hub-skill --vet        # Force local re-scan
ipman install local.ip.yaml --no-vet # Skip local scan

Security Log

Blocked and warned installs are logged to ~/.ipman/security.log:

2026-03-14T10:30:00Z BLOCKED sketchy-tool source=iphub risk=EXTREME reason="credential harvesting detected"

Disable via config:

security:
  log_enabled: false

Report Suspicious Skills

ipman hub report sketchy-tool --reason "Accesses ~/.ssh without clear purpose"

Report counts feed back into risk scoring and are displayed in ipman hub info.