latent-sre/homelab-agents/security-audit

Audit the lab the way an attacker reads it: not "is it well-kept" but "can someone get in, move, or take." Every finding is evidence-cited and carries the path an attacker would actually walk — sde-agents:lab-audit owns the hygiene sweep; this skill owns the adversary's question.

All checks are read-only. disallowed-tools removes Write and Edit while this skill is active, but Bash can still mutate (redirects, docker rm), so the mandate is still yours: inspection commands only — every fix routes to sde-agents:homelab-engineer, and vulnerability findings feed sde-agents:upgrade-campaign's priority order rather than becoming ad-hoc patches. The read-only-ness here is cooperative, not enforced (the reviewer's Bash guard keys on guarded agent identities, not skills). Fan the checks out in parallel (per zone or per check area) rather than sweeping serially.

Two rules with no exceptions:

• A finding carries an attack path or gets downgraded. A pattern match with no reachable

route from an attacker position is a P2/P3 note, not a P0 — say what position the attacker needs, what they cross, and what they reach, or lower the severity and say why.

• Triage suspicion; escalate corroborated compromise. An unfamiliar key, process, or container

is a lead. Make a bounded, nonmutating provenance check against deployment records, key fingerprints and ownership, package/image identity, and relevant access logs. Select only needed fields and keep credentials out of output. An explained, authorized artifact lets the sweep continue. An unresolved lead retains its evidence and missing provenance in the report; uncertainty is not proof of a breach. Corroborated unauthorized access, execution, exfiltration, or tampering ends the ordinary sweep: preserve available evidence and hand the incident to the operator. Do not clean up, restart, rebuild, rotate credentials, or isolate hosts as an audit step. Containment and recovery use sde-agents:lab-incident under sde-agents:homelab-engineer's authority and actual host controls.

Checks (run what applies; name what you skipped in the denominator)

The seven checks — trust zones and reachability, authentication on exposed services, management planes, credentials, secrets posture, vulnerability triage, personal-data paths — live with their command-level detail in `references/checks.md`; read it before sweeping. The secrets check has its own deep-dive at `references/secrets.md`, loaded when that row trips.

Output

Open with the coverage denominator — zones and checks swept vs. skipped, with why — then findings ranked [P0]–[P3], each with its evidence (command + output — for secrets and credentials, names and paths only, never values: a report that quotes a secret is itself a leak), its attack path (position → crossing → reach), and the one-line fix class. Grade severity by the unauthorized capability or data reached and the prerequisites; intended public content and permitted, authenticated access are not findings by themselves. End with the top three things to fix this weekend, then emit the findings-ledger rows in sde-agents:lab-audit's table format for the operator to append to the lab repo's ledger — this skill holds no write tools, so the emitted block IS the ledger entry.