yeaboi

Agent Security Beta

An agent is only as safe as the setup around it. This mode audits that setup — the permission rules, the MCP servers, the secrets within reach, and the risky commands already run — and scores the posture.

Beta. The Agents family is new, and every number it shows is computed from telemetry your agents happen to leave behind. Treat it as an indicator, not an audit.

Every check is a deterministic pattern scan. It is an indicator, not a security audit: it will tell you that an allow rule auto-approves Bash, not whether that is wrong for your project.

What it checks

  • Settings — permission modes that bypass prompting, and wildcard allow rules that auto-approve broad tool classes.
  • MCP servers — an inventory of what is wired into your agents, since every server is reachable surface.
  • Secrets exposure — credential-shaped strings sitting where an agent session can read them.
  • Risky commands — shapes already present in session history: piping a download straight into a shell, decoding base64 into a shell, rm -rf /, permission-skip flags, and sudo.

Findings are ranked by severity — critical, high, medium, info — and rolled into an overall posture score.

What a finding contains

A finding carries a pattern label, a file path, and a line number where that is meaningful. It never carries the secret it matched, the command it appeared in, or any prompt or code from a session. The only value a finding may quote is from your own configuration — the permission mode, the allow rule — because a finding that cannot name the rule cannot be acted on, and your settings file is not session content.

Unreadable or malformed config produces a note-level finding rather than an error: a security page that died on one corrupt JSON file would hide every other finding on it.

Deep scans

An ordinary scan only re-reads transcripts that changed since last time. --deep forgets those cursors first, so every transcript is scanned again — slower, and the right choice after adding a pattern or when you want a clean baseline.

yeaboi agents security
yeaboi agents security --deep
yeaboi agents security --format json

Over MCP: agents_security_scan and agents_security_history.