security-audit

name	security-audit
description	Multi-phase security audit skill. Scans agentic config surfaces and application source for OWASP Top 10 vulnerabilities, STRIDE threats, and config hygiene issues. Produces a scored, graded report persisted to the configured security docs directory (default: docs/security).
argument-hint	[mode:report|mode:fix] [scope:full|scope:config|scope:owasp|scope:stride|scope:<path>]

Security Audit

Performs a comprehensive security audit of the workspace across agentic configuration surfaces and application source code. Produces a scored, graded security report.

Invocation

Invoke security-audit [mode:report|mode:fix] [scope:full|scope:config|scope:owasp|scope:stride|scope:<path>]

Defaults: mode:report, scope:full

mode:report — Analyze and report findings. Never modify files. mode:fix — Apply config-tier fixes only (Tier 1 deterministic fixes). Never auto-fix OWASP or STRIDE findings.

scope:full — Run all phases (default) scope:config — Run Phase 1 (discovery) and Phases 2–3 (config audit only) scope:owasp — Run Phase 1 (discovery) and Phase 4 (OWASP source scan) scope:stride — Run Phase 1 (discovery) and Phase 5 (STRIDE threat model) scope:<path> — Run full audit (all phases) scoped to the specified path; config phases apply only if the path intersects config surfaces

Subagent Depth Constraint

This is a leaf executor. No subagent spawning. Maximum depth: 0.

Variables

Variable	Purpose
.github/**,.vscode/**,.copilot/**	Glob patterns for agentic config files (e.g., .github/**, .vscode/**)
src/**	Application source file patterns (e.g., src/**/*.rs)
Rust	Primary language for pattern selection
docs/security	Output directory for persisted reports (default: docs/security)
`	Rule
---	---
Terminal auto-approve patterns	.vscode/settings.json chat.tools.terminal.autoApprove must not include destructive commands
Hardcoded credentials	No API keys, tokens, or passwords in .github/, .vscode/, or instruction files
Overly permissive allow-lists	.github/copilot-instructions.md and agent definitions must not blanket-approve dangerous operations
Unpinned secrets	Workflow secrets must use environment-scoped access, not repo-wide
Debug endpoints	No debug or diagnostic endpoints exposed without authentication
- **Injection**: Check for unsafe blocks, unchecked deserialization (serde_json::from_str` on untrusted input without validation), SQL injection via raw SurrealDB queries

• Broken Auth: Verify IPC transport validates caller identity, check lockfile permissions
• Sensitive Data Exposure: Check for credentials or tokens in log output, verify tracing filters exclude sensitive fields
• Security Misconfiguration: Verify daemon binds to 127.0.0.1 only, check feature flags do not accidentally expose debug endpoints
• Deserialization: Verify all MCP tool input is validated before processing, check for unbounded deserialization of user-controlled data` | Language-specific OWASP detection patterns |

Workflow

Phase 1: Discovery

Skip condition: Always runs. For focused scopes, discovery narrows to the relevant surfaces:

• scope:config — enumerate only agentic config surfaces
• scope:owasp or scope:stride — enumerate only application source entry points
• scope:<path> — enumerate only files under the specified path
• scope:full — enumerate all surfaces

1. Enumerate agentic config surfaces matching .github/**,.vscode/**,.copilot/**
2. Enumerate application source files matching src/**
3. Identify the primary language (Rust) and select corresponding OWASP patterns
4. Record the audit scope, file counts, and entry points found

Phase 2: Config Tier 1 — Deterministic Config Scan

Skip condition: Skip unless scope:full, scope:config, or scope:<path> where the scoped path intersects config surfaces.

Apply deterministic regex checks to config surfaces found in Phase 1. Findings in this tier are eligible for mode:fix auto-remediation.

Rules from | Rule | Check | Scope | |---|---|---| | Terminal auto-approve patterns | .vscode/settings.json chat.tools.terminal.autoApprovemust not include destructive commands | VS Code | | Hardcoded credentials | No API keys, tokens, or passwords in.github/, .vscode/, or instruction files | All | | Overly permissive allow-lists | .github/copilot-instructions.md and agent definitions must not blanket-approve dangerous operations | GitHub Copilot | | Unpinned secrets | Workflow secrets must use environment-scoped access, not repo-wide | GitHub Actions | | Debug endpoints | No debug or diagnostic endpoints exposed without authentication | Runtime |:

• Hardcoded credential patterns (passwords, tokens, keys) in config files
• Overly permissive tool allow-lists (e.g., always: true on destructive terminal commands)
• Missing input validation rules in agent instruction files
• Exposed secrets in environment variable defaults
• Missing branch protection or CI security gate references

For each match: record file, line number, matched pattern, severity, and fix action.

Phase 3: Config Tier 2 — LLM-Assessed Config Rules

Skip condition: Skip unless scope:full, scope:config, or scope:<path> where the scoped path intersects config surfaces.

Evaluate config files that require semantic interpretation rather than regex matching:

• Agent instruction files that grant broad tool permissions without scope constraints
• CI workflows that allow arbitrary code execution from pull requests
• Dependency configuration files with unpinned versions in security-sensitive packages
• Agentic tool configurations that bypass the workspace's own safety policies

Apply judgment: record findings with reasoning, not just pattern matches. These findings are advisory in mode:fix — they require human review before action.

Phase 4: OWASP Top 10 Source Scan

Skip condition: Skip unless scope:full, scope:owasp, or scope:<path>.

Scan source files matching src/** (or the specified path) using - **Injection**: Check for unsafe blocks, unchecked deserialization (serde_json::from_str` on untrusted input without validation), SQL injection via raw SurrealDB queries

• Broken Auth: Verify IPC transport validates caller identity, check lockfile permissions
• Sensitive Data Exposure: Check for credentials or tokens in log output, verify tracing filters exclude sensitive fields
• Security Misconfiguration: Verify daemon binds to 127.0.0.1 only, check feature flags do not accidentally expose debug endpoints
• Deserialization: Verify all MCP tool input is validated before processing, check for unbounded deserialization of user-controlled data`:

Category	What to look for
A01: Broken Access Control	Authorization checks missing at data-object layer, not just route layer
A02: Cryptographic Failures	Sensitive data stored or transmitted without encryption; weak hash algorithms
A03: Injection	String-built queries, unparameterized SQL, unescaped template values, shell command construction from inputs
A04: Insecure Design	Missing rate limiting on auth endpoints, missing CSRF protection on state-changing routes
A05: Security Misconfiguration	Default credentials, debug endpoints, verbose error responses in production paths
A06: Vulnerable Components	Package manifests with known-vulnerable pinned versions (if detectable without network)
A07: Auth Failures	Session token generation entropy, missing invalidation on logout, fixation vulnerabilities
A08: Software Integrity	Dynamic import of unvalidated paths, deserialization of untrusted data
A09: Logging Failures	Sensitive values (passwords, tokens, PII) interpolated into log statements
A10: SSRF	URL parameters passed to HTTP clients without allowlist validation

OWASP findings are report-only. Do not auto-fix source code even in mode:fix.

Phase 5: STRIDE Threat Model

Skip condition: Skip unless scope:full, scope:stride, or scope:<path>.

Apply a lightweight STRIDE analysis to the system's trust model:

Threat	Question
Spoofing	Can an attacker impersonate a user, service, or agent?
Tampering	Can inputs or stored data be modified without detection?
Repudiation	Can actors deny their actions? Is sufficient audit logging present?
Information Disclosure	Can sensitive data be accessed by unauthorized parties?
Denial of Service	Can resource exhaustion or crash loops be triggered from external inputs?
Elevation of Privilege	Can an actor gain higher privileges than intended?

For each STRIDE category, identify the highest-risk finding from the workspace scan. Report STRIDE findings as advisory — they describe design-level risks, not specific exploitable code patterns. STRIDE findings do not require an exact file:line location; provide an "evidence anchor" (the specific system component, data flow, or design decision that introduces the risk) instead.

Phase 6: Score and Grade

Compute the security score using this deduction model applied to a 100-point baseline:

Severity	Deduction
Critical (P0)	-15 per finding
High (P1)	-10 per finding
Medium (P2)	-5 per finding
Low (P3)	-2 per finding

N/A semantics: If a category does not apply to the workspace (e.g., no web routes → A10 SSRF is N/A), exclude it from scoring rather than scoring it as 0 findings.

Final grade:

Score	Grade
90–100	A — Low risk
75–89	B — Moderate risk, remediation advised
60–74	C — Elevated risk, remediation required before release
< 60	F — High risk, block release until critical/high findings resolved

Phase 7: Output

Produce a structured report with:

1. Executive Summary — score, grade, finding counts by severity, top 3 risks, audit scope and date
2. Config Audit Results (Phases 2–3) — findings grouped by rule category
3. OWASP Scan Results (Phase 4) — findings grouped by A01–A10 category
4. STRIDE Analysis (Phase 5) — one entry per STRIDE category with the highest-risk finding and its rationale
5. Remediation Priority List — ordered action list: P0 → P1 → P2 → P3, each with location, description, and remediation steps
6. Applied Fixes (mode:fix only) — list of Tier 1 config fixes applied with before/after diff

Phase 8: Persist

Save the report to docs/security/YYYY-MM-DD-HH-MM-security-audit.md where YYYY-MM-DD is today's date and HH-MM is the current time (24-hour). The time component prevents clobbering earlier runs on the same day when the skill is invoked multiple times (different scopes or modes).

If docs/security/ does not exist, create it before writing.

Report the saved path on completion.

Mode Rules

mode:report (default):

• Read and search only
• Never modify source or config files
• Report all findings with location and remediation guidance

mode:fix:

• Apply only Phase 2 (Tier 1 deterministic) config fixes
• Config fixes must be file edits with clear before/after rationale
• Never modify application source code
• Never apply Phase 3, 4, or 5 findings as auto-fixes
• After applying fixes, re-run Phase 2 to confirm the fix resolved the finding

Quality Criteria

• Every phase that is in scope must run before the report is considered complete
• Skipped phases must be noted in the executive summary with the skip reason
• All findings include a specific file:line location, except STRIDE findings which require an evidence anchor (component, data flow, or design decision) instead of an exact line number
• The report is persisted to docs/security/ before the audit is marked complete

Model Routing

This skill operates at Tier 2 (Standard) for config audit phases and Tier 3 (Frontier) for OWASP and STRIDE analysis phases. If only a single model is available, use it for all phases.

Generated by autoharness | Template: security-audit/SKILL.md.tmpl