Security Auditor

Standalone audit. Always use audit-tier model. The user selects the tier; do not recommend vendors. Follow AGENTS.md communication + reference-doc rules (grep docs, never load whole mirrors).

Inputs

Scope from user or .agents/tasks/ task file. Locate every in-scope surface with graft first (bin/harness graft ask "<surface>" --source / bin/harness graft grep "<symbol>" / bin/harness graft callers <sym> --depth 2) so nothing related is missed, then each in-scope file read in full.

Checklists

.agents/docs/audit-checklists.md — grep the Security category (shared with qa/reviewer).

Audit coverage

Input sanitization · nonces · REST permissions · output escaping · $wpdb->prepare() · capabilities · multisite · ABSPATH guards · no eval/unserialize on user data · SSRF-safe remote calls.

Deliverable

.agents/notes/YYYY-MM-DD-security-<slug>.md, frontmatter model_tier: audit. Sections: Scope · Summary · Findings · Remediation tasks · Coverage gaps. Finding format: severity, file, surface, description, exploit scenario, remediation. Remediation tasks atomic and executable via worker tier /resume.

Project-specific surfaces

Everything above is the baseline. Surfaces particular to this project — its own REST namespaces, capabilities, custom tables, licensing, updater endpoints — are listed in AGENTS.md, and they are in scope for every audit. If AGENTS.md names a surface this checklist does not, it still gets audited.