atv-security

# /atv-security — Unified Security Auditor Scan your project for security issues across two surfaces: 1. **Agentic configuration** — `.github/`, `.vscode/` configs (33 rules adapted from [AgentShield](https://github.com/affaan-m/agentshield)). 2. **Application source code** — OWASP Top 10 (2021) static checks + STRIDE threat model. > **Heritage:** This skill absorbs the former `/cso` skill. Old `/cso` triggers still route here. The on-disk security report file keeps the legacy `<!-- cso -->` marker block so existing reports stay structurally compatible. **5 config categories:** Secrets · Permissions · Hooks · MCP Servers · Agents & Skills ## Arguments <args> #$ARGUMENTS </args> **Argument grammar:** Two independent axes, parsed in order. 1. **Mode** — if any token in `$ARGUMENTS` matches `fix` (case-insensitive), mode = `fix`. Otherwise mode = `report` (default). 2. **Scope** — examine the remaining tokens (after stripping `fix`): - Token `config` → scope = `config` (config audit only; skip OWASP/STRIDE) - Token `owasp` → scope = `owasp` (OWASP scan only; skip config + STRIDE) - Token `stride` → scope = `stride` (STRIDE only; skip config + OWASP) - Token that looks like a file/directory path (contains `/` or `\` or matches an existing path) → scope = `<path>` (run OWASP/STRIDE narrowed to that path; skip config) - No remaining tokens, or token is `full` → scope = `full` (default; run everything available) **Examples:** - `/atv-security` → `mode=report scope=full` - `/atv-security fix` → `mode=fix scope=full` - `/atv-security config fix` or `/atv-security fix config` → `mode=fix scope=config` - `/atv-security owasp` → `mode=report scope=owasp` - `/atv-security src/api/` → `mode=report scope=src/api/` ## Execution Flow ``` Phase 1: Discovery → Detect config surfaces + app source stack Phase 2: Tier 1 Config → Deterministic regex scan of .github/, .vscode/ (skip if scope ∉ {full, config}) Phase 3: Tier 2 Config → LLM-assessed config rules (skip if scope ∉ {full, config}) Phase 4: OWASP Top 10 → Application source code scan (skip if no source OR scope ∉ {full, owasp, <path>}) Phase 5: STRIDE → Threat model the application (skip if no source OR scope ∉ {full, stride, <path>}) Phase 6: Score & Grade → Per-surface grades with N/A semantics Phase 7: Output → Combined report (config + OWASP + STRIDE) Phase 8: Persist → Upsert into docs/security/YYYY-MM-DD-security-report.md (both marker blocks) Phase 9: Fix → Opt-in safe fixes for auto-fixable config rules (only when mode=fix) ``` --- ## Phase 1: Discovery Use `file_search` and `list_dir` to detect both surfaces in parallel. ### 1a. ATV configuration surfaces | Surface | File Pattern | Category | |---------|-------------|----------| | Instructions | `.github/copilot-instructions.md` | Agents & Skills | | MCP Config | `.github/copilot-mcp-config.json` | MCP Servers | | Skills | `.github/skills/**/*.md` | Agents & Skills | | Agents | `.github/agents/**/*.agent.md` | Agents & Skills | | Hooks | `.github/hooks/copilot-hooks.json` + `.github/hooks/scripts/**` | Hooks | | Setup Steps | `.github/copilot-setup-steps.yml` | Hooks | | VS Code | `.vscode/settings.json`, `.vscode/extensions.json` | Permissions | Set `hasConfig = true` if any of the above are found. ### 1b. Application source stack | Signal | Stack | Key files to scan | |--------|-------|-------------------| | `package.json`, `*.ts`, `*.js` | Node.js / TypeScript | `src/**`, `routes/**`, `api/**`, `pages/**` | | `requirements.txt`, `*.py` | Python | `app/**`, `src/**`, `views/**`, `api/**` | | `Gemfile`, `*.rb` | Ruby / Rails | `app/**`, `config/**`, `db/**` | | `go.mod`, `*.go` | Go | `**/*.go` | | `*.cs`, `*.csproj` | .NET | `**/*.cs`, `Controllers/**` | | `pom.xml`, `*.java` | Java | `src/**/*.java` | Set `hasSource = true` if any application source files are found. If scope is a path, narrow source detection to that path. Record: list of discovered config files, detected stack, total source files found, entry points identified. ### 1c. Bail rule If `!hasConfig && !hasSource`: report "No ATV configuration or application source detected. Run `npx atv-starterkit init` to scaffold an agentic environment, or run `/atv-security` from a project directory." and stop. If `scope=config` and `!hasConfig`: report "Scope `config` requested but no `.github/` directory found." and stop. If `scope ∈ {owasp, stride, <path>}` and `!hasSource`: report "Scope requires application source, but none found." and stop. --- ## Phase 2: Tier 1 Config Scan — Deterministic Detection > Run only when `hasConfig && scope ∈ {full, config}`. Run `grep_search` with `isRegexp: true` for each pattern below. For each match, record a finding with the specified fields. ### Secrets Rules | Rule | Pattern | Scope | Severity | Fix | |------|---------|-------|----------|-----| | SEC-01 | `sk-ant-[a-zA-Z0-9]{20,}` | All `.github/**`, `.vscode/**` | 🔴 critical | Replace with `${ANTHROPIC_API_KEY}` env var reference | | SEC-02 | `sk-proj-[a-zA-Z0-9]{20,}` | All `.github/**`, `.vscode/**` | 🔴 critical | Replace with `${OPENAI_API_KEY}` env var reference | | SEC-03 | `AKIA[0-9A-Z]{16}` | All `.github/**`, `.vscode/**` | 🔴 critical | Replace with `${AWS_ACCESS_KEY_ID}` env var reference | For rules whose regex patterns require alternation, use the entries below instead of markdown table rows so the raw `|` characters remain valid for `isRegexp: true`: - **SEC-04** - **Pattern:** `(?:ghp_[a-zA-Z0-9]{36}|github_pat_[a-zA-Z0-9]{22,})` - **Scope:** All `.github/**`, `.vscode/**` - **Severity:** 🔴 critical - **Fix:** Replace with `${GITHUB_TOKEN}` env var reference - **SEC-05** - **Pattern:** `(?:Bearer [a-zA-Z0-9_\-\.]{20,}|mongodb(\+srv)?://[^\s]+|postgres(ql)?://[^\s]+|mysql://[^\s]+|redis://[^\s]+)` - **Scope:** All `.github/**`, `.vscode/**` - **Severity:** 🟡 high - **Fix:** Replace with `${ENV_VAR}` reference appropriate to the service ### MCP Server Rules (grep-detectable) | Rule | Pattern | Scope | Severity | Fix | |------|---------|-------|----------|-----| | MCP-02 | `"tools"\s*:\s*\["?\*"?\]` | `.github/copilot-mcp-config.json` | 🟡 high | Scope to specific tools needed: `["tool1", "tool2"]` | | MCP-03 | `autoApprove` | `.github/copilot-mcp-config.json` | 🟢 medium | Remove autoApprove or restrict to safe read-only tools | - **MCP-04** - **Pattern:** `(?:sk-ant-|sk-proj-|AKIA|ghp_|Bearer )` - **Scope:** `.github/copilot-mcp-config.json` env sections - **Severity:** 🔴 critical - **Fix:** Use `${input:VAR}` or `${ENV_VAR}` references ### Hook Rules (grep-detectable) - **HOOK-01** - **Pattern:** `(?:curl.*\$\{|wget.*\$\{|eval.*\$\{)` - **Scope:** `.github/hooks/scripts/**` - **Severity:** 🟡 high - **Fix:** Validate/sanitize variables before use in network/eval commands - **HOOK-02** - **Pattern:** `(?:curl\s+-X\s+POST.*\$|wget\s+--post)` - **Scope:** `.github/hooks/scripts/**` - **Severity:** 🔴 critical - **Fix:** Remove data exfiltration patterns or restrict to known-safe URLs | Rule | Pattern | Scope | Severity | Fix | |------|---------|-------|----------|-----| | HOOK-03a | `2>/dev/null` | `.github/hooks/scripts/**` | 🟢 medium | Log errors instead of suppressing them silently | | HOOK-03b | `\|\| true$` | `.github/hooks/scripts/**` | 🟢 medium | Log errors instead of suppressing them silently | | HOOK-03c | `\|\| exit 0$` | `.github/hooks/scripts/**` | 🟢 medium | Log errors instead of suppressing them silently | ### Agent & Skill Rules (grep-detectable) | Rule | Pattern | Scope | Severity | Fix | |------|---------|-------|----------|-----| | AGENT-01 | `[\u200B\u200C\u200D\uFEFF]` | `.github/skills/**`, `.github/agents/**`, `.github/copilot-instructions.md` | 🔴 critical | Remove zero-width characters — likely hidden instruction injection | | AGENT-02 | `[A-Za-z0-9+/]{80,}={0,2}` | `.github/skills/**`, `.github/agents/**` | 🟢 medium | Decode and inspect — may contain hidden instructions. Ignore if preceded by `sha256:`, `data:`, or `http` | ### Permission Rules (grep-detectable) | Rule | Pattern | Scope | Severity | Fix | |------|---------|-------|----------|-----| | PERM-01 | `security\.workspace\.trust\.enabled"?\s*:\s*false\|chat\.tools\.autoApprove"?\s*:\s*true` | `.vscode/settings.json` | 🟢 medium | Enable workspace trust; disable agent-tool auto-approval | **Execution:** For each rule, run `grep_search` with the pattern and `includePattern` matching the scope. Record every match as a finding with: rule ID, category, severity, title, file path, matched evidence (truncated to 100 chars), and fix suggestion. ---