agentaudit 3.13.4 → 3.13.6

package/README.md

@@ -77,18 +77,18 @@ agentaudit lookup fastmcp
 
 **Example output:**
 ```
-  ⛨ AgentAudit v3.12.9  │  my-scanner #3 · 280pts · 19 audits
+  ◆ AgentAudit v3.13.4  │  my-scanner · #3 · 280pts · 19 audits
 
   Discovering MCP servers in your AI editors...
 
 •  Scanning Cursor  ~/.cursor/mcp.json    found 3 servers
 
 ├──  tool   supabase-mcp              ✔ ok
-│   SAFE  Risk 0  https://agentaudit.dev/skills/supabase-mcp
+│   SAFE  Risk 0  https://agentaudit.dev/packages/supabase-mcp
 ├──  tool   browser-tools-mcp         ✔ ok
 │   ⚠ not audited  Run: agentaudit audit https://github.com/nichochar/browser-tools-mcp
 └──  tool   filesystem                ✔ ok
-│   SAFE  Risk 0  https://agentaudit.dev/skills/filesystem
+│   SAFE  Risk 0  https://agentaudit.dev/packages/filesystem
 
   Looking for general package scanning? Try `pip audit` or `npm audit`.
 ```
@@ -210,7 +210,11 @@ Then ask your agent: *"Check which MCP servers I have installed and audit any un
 | `agentaudit scan <url>` | Quick regex-based static scan (~2s) | `agentaudit scan https://github.com/owner/repo` |
 | `agentaudit scan <url> --deep` | Deep audit (same as `audit`) | `agentaudit scan https://github.com/owner/repo --deep` |
 | `agentaudit audit <url>` | Deep LLM-powered 3-pass audit (~30s) | `agentaudit audit https://github.com/owner/repo` |
+| `agentaudit audit <url> --verify` | Audit + adversarial verification pass (reduces false positives) | `agentaudit audit <url> --verify self` |
+| `agentaudit audit <url> --remote` | Server-side scan via agentaudit.dev (no LLM key needed, 3/day free) | `agentaudit audit <url> --remote` |
+| `agentaudit consensus <name>` | Cross-model consensus view for a package | `agentaudit consensus supabase-mcp` |
 | `agentaudit lookup <name>` | Look up package in trust registry | `agentaudit lookup fastmcp` |
+| `agentaudit history` | Show local audit history | `agentaudit history` |
 
 ### Community
 
@@ -238,6 +242,10 @@ Then ask your agent: *"Check which MCP servers I have installed and audit any un
 | `--quiet` / `-q` | Suppress banner and decorative output |
 | `--no-color` | Disable ANSI colors (also respects `NO_COLOR` env var) |
 | `--model <name>` | Override LLM model for this run |
+| `--models <a,b,c>` | Multi-model audit (parallel calls, consensus comparison) |
+| `--verify <mode>` | Adversarial verification: `self` (same model), `cross` (different model), or `<model-name>` |
+| `--no-verify` | Skip verification even if configured |
+| `--remote` | Use agentaudit.dev server for scan (no local LLM key needed) |
 | `--no-upload` | Skip uploading report to registry |
 | `--export` | Export audit payload as markdown |
 | `--debug` | Show raw LLM response on parse errors |
@@ -279,6 +287,9 @@ When running as an MCP server, AgentAudit exposes the following tools to your AI
 | `check_registry` | Look up a package in the trust registry |
 | `submit_report` | Upload audit findings to the registry |
 | `discover_servers` | Find MCP servers in local editor configs |
+| `consensus_analysis` | Cross-model consensus view for a package |
+| `search_packages` | Search packages in the registry by name, ASF-ID, or hash |
+| `scan_tool_poisoning` | Detect tool poisoning in MCP tool descriptions |
 
 ### Workflow
 
@@ -383,6 +394,34 @@ The deep audit (`agentaudit audit`) uses a structured 3-phase LLM analysis — n
 
 This architecture achieved **0% false positives** on our 11-package test set, down from 42% in v2.
 
+### Adversarial Verification Pass (v3.13+)
+
+After the 3-pass audit, an optional **verification pass** re-examines each finding against the actual source code:
+
+```bash
+agentaudit audit https://github.com/owner/repo --verify self
+```
+
+Each finding goes through a 5-point checklist:
+1. **Code Existence** — Does the cited code actually exist in the file?
+2. **Context Accuracy** — Is the code used in the way described?
+3. **Execution Model** — Can an attacker actually trigger this?
+4. **Severity Calibration** — Is the severity appropriate?
+5. **Fabrication Check** — Are there hallucinated details?
+
+Verdicts: `verified` (confirmed real), `demoted` (severity reduced), `rejected` (false positive removed).
+
+### Model Accuracy (Real-World Data)
+
+We benchmarked multiple LLMs on the **Top 20 most popular MCP servers** (62+ reports):
+
+| Model | Findings on Top 20 | Precision | Assessment |
+|-------|-------------------|-----------|------------|
+| **Claude Opus 4.6** | 0 findings (all clean) | N/A | Very conservative — ideal for avoiding false positives |
+| **Gemini 2.5 Flash** | Many findings | ~30% strict | High false positive rate — not recommended for production audits |
+
+> **Key insight:** Model choice dramatically affects audit quality. We recommend Claude Opus 4 or Claude Sonnet 4 for production audits. Use `--models` to run multiple models and compare results via `consensus`.
+
 ---
 
 ## 🔄 CI/CD Integration
@@ -450,11 +489,12 @@ AgentAudit includes a full-screen interactive dashboard and standalone community
 agentaudit dashboard    # or: agentaudit dash
 ```
 
-5-tab TUI with keyboard navigation (←→ tabs, ↑↓ scroll, 1-5 jump, q quit):
+5-tab TUI with keyboard navigation (←→ tabs, ↑↓ scroll, 1-5 jump, q quit).
+Overview tab includes **interactive Quick Actions** — select and launch audits, consensus views, or remote scans directly from the dashboard:
 
 | Tab | Content |
 |-----|---------|
-| **[1] Overview** | Your profile (rank, points, audits, severity breakdown) + registry stats |
+| **[1] Overview** | Your profile + registry stats + interactive Quick Actions (press a/v/r/c or Enter) |
 | **[2] Leaderboard** | Top contributors with medal rankings and bar charts |
 | **[3] Benchmark** | LLM model audit performance comparison |
 | **[4] Activity** | Your recent audits and findings |

package/cli.mjs

@@ -64,7 +64,7 @@ const LLM_PROVIDERS = [
   { key: 'DEEPSEEK_API_KEY',    name: 'DeepSeek',              provider: 'deepseek',    type: 'openai',    model: 'deepseek-chat',             url: 'https://api.deepseek.com/v1/chat/completions' },
   { key: 'MISTRAL_API_KEY',     name: 'Mistral',               provider: 'mistral',     type: 'openai',    model: 'mistral-large-latest',      url: 'https://api.mistral.ai/v1/chat/completions' },
   { key: 'GROQ_API_KEY',        name: 'Groq',                  provider: 'groq',        type: 'openai',    model: 'llama-3.3-70b-versatile',   url: 'https://api.groq.com/openai/v1/chat/completions' },
-  { key: 'XAI_API_KEY',         name: 'xAI (Grok)',            provider: 'xai',         type: 'openai',    model: 'grok-3',                    url: 'https://api.x.ai/v1/chat/completions' },
+  { key: 'XAI_API_KEY',         name: 'xAI (Grok)',            provider: 'xai',         type: 'openai',    model: 'grok-4',                    url: 'https://api.x.ai/v1/chat/completions' },
   { key: 'TOGETHER_API_KEY',    name: 'Together AI',           provider: 'together',    type: 'openai',    model: 'meta-llama/Llama-3.3-70B-Instruct-Turbo', url: 'https://api.together.xyz/v1/chat/completions' },
   { key: 'FIREWORKS_API_KEY',   name: 'Fireworks AI',          provider: 'fireworks',   type: 'openai',    model: 'accounts/fireworks/models/llama-v3p3-70b-instruct', url: 'https://api.fireworks.ai/inference/v1/chat/completions' },
   { key: 'CEREBRAS_API_KEY',    name: 'Cerebras',              provider: 'cerebras',    type: 'openai',    model: 'llama-3.3-70b',             url: 'https://api.cerebras.ai/v1/chat/completions' },
@@ -78,15 +78,16 @@ const LLM_PROVIDERS = [
 const PROVIDER_MODELS = {
   anthropic: [
     { label: 'claude-sonnet-4-20250514', sublabel: 'fast + smart (default)', value: 'claude-sonnet-4-20250514' },
-    { label: 'claude-opus-4-20250514',   sublabel: 'most capable',           value: 'claude-opus-4-20250514' },
+    { label: 'claude-opus-4-20250514',   sublabel: 'best precision (recommended for audits)', value: 'claude-opus-4-20250514' },
   ],
   openai: [
     { label: 'gpt-4o',  sublabel: 'fast multimodal (default)', value: 'gpt-4o' },
-    { label: 'gpt-4.1', sublabel: 'latest',                    value: 'gpt-4.1' },
+    { label: 'gpt-4.1', sublabel: 'large context (low recall on audits)', value: 'gpt-4.1' },
   ],
   google: [
     { label: 'gemini-2.5-flash', sublabel: 'fast + cheap (default)', value: 'gemini-2.5-flash' },
-    { label: 'gemini-2.5-pro',   sublabel: 'most capable',          value: 'gemini-2.5-pro' },
+    { label: 'gemini-2.5-pro',   sublabel: 'strong reasoning',      value: 'gemini-2.5-pro' },
+    { label: 'gemini-3.1-pro',   sublabel: 'best detection (recommended for audits)', value: 'gemini-3.1-pro' },
   ],
   deepseek: [
     { label: 'deepseek-chat', sublabel: 'cost-effective (default)', value: 'deepseek-chat' },
@@ -98,7 +99,8 @@ const PROVIDER_MODELS = {
     { label: 'llama-3.3-70b-versatile', sublabel: 'ultra-fast (default)', value: 'llama-3.3-70b-versatile' },
   ],
   xai: [
-    { label: 'grok-3', sublabel: 'real-time knowledge (default)', value: 'grok-3' },
+    { label: 'grok-4', sublabel: 'best detection (default, recommended)', value: 'grok-4' },
+    { label: 'grok-3', sublabel: 'faster, lower cost',                    value: 'grok-3' },
   ],
   together: [
     { label: 'meta-llama/Llama-3.3-70B-Instruct-Turbo', sublabel: 'open source (default)', value: 'meta-llama/Llama-3.3-70B-Instruct-Turbo' },
@@ -267,6 +269,14 @@ function resolveProvider() {
 }
 
 function resolveModel(modelName) {
+  // Shorthand aliases for recommended models
+  const aliases = {
+    'opus': 'claude-opus-4-20250514',
+    'sonnet': 'claude-sonnet-4-20250514',
+    'gemini-3.1-pro': 'google/gemini-3.1-pro-preview',
+    'gemini-3.1-flash': 'google/gemini-3.1-flash-preview',
+  };
+  if (aliases[modelName.toLowerCase()]) modelName = aliases[modelName.toLowerCase()];
   // model with '/' → OpenRouter
   if (modelName.includes('/')) {
     const p = LLM_PROVIDERS.find(p => p.provider === 'openrouter' && process.env[p.key]);
@@ -2812,18 +2822,26 @@ Decision rules: code_exists=false→REJECTED; code_matches_description=false→R
 
 // Known context window sizes (input tokens) for common models
 const MODEL_CONTEXT_LIMITS = {
+  'claude-sonnet-4-6': 200000, 'claude-opus-4-6': 200000,
   'claude-sonnet-4': 200000, 'claude-opus-4': 200000, 'claude-haiku-4': 200000,
   'claude-3.5-sonnet': 200000, 'claude-3-haiku': 200000,
+  'gpt-4.1': 1047576, 'gpt-4.1-mini': 1047576, 'gpt-4.1-nano': 1047576,
   'gpt-4o': 128000, 'gpt-4o-mini': 128000, 'gpt-4-turbo': 128000, 'gpt-4': 8192,
+  'gemini-3.1-pro': 1048576, 'gemini-3.1-flash': 1048576,
   'gemini-2.5-flash': 1048576, 'gemini-2.5-pro': 1048576, 'gemini-2.0-flash': 1048576,
+  'grok-4': 256000, 'grok-3': 131072,
   'deepseek-chat': 64000, 'deepseek-reasoner': 64000,
   'mistral-large': 128000, 'mistral-small': 32000,
 };
 
 function estimateTokens(text) { return Math.ceil(text.length / 3.5); }
 
+// Sorted keys: longest first so "gpt-4.1" matches before "gpt-4", "claude-sonnet-4-6" before "claude-sonnet-4"
+const MODEL_LIMIT_KEYS = Object.keys(MODEL_CONTEXT_LIMITS).sort((a, b) => b.length - a.length);
+
 function checkContextLimit(model, systemPrompt, userMessage) {
-  const modelKey = Object.keys(MODEL_CONTEXT_LIMITS).find(k => model.toLowerCase().includes(k.toLowerCase()));
+  const stripped = model.replace(/^(anthropic|openai|google|openrouter|meta-llama|mistralai)\//i, '').toLowerCase();
+  const modelKey = MODEL_LIMIT_KEYS.find(k => stripped.includes(k.toLowerCase()));
   if (!modelKey) return null; // unknown model, skip check
   const limit = MODEL_CONTEXT_LIMITS[modelKey];
   const estimated = estimateTokens(systemPrompt) + estimateTokens(userMessage);
@@ -5285,7 +5303,7 @@ async function main() {
       `  agentaudit audit https://github.com/owner/repo --verify cross`,
       `  agentaudit audit https://github.com/owner/repo --remote`,
       `  agentaudit audit https://github.com/owner/repo --model gpt-4o`,
-      `  agentaudit audit https://github.com/owner/repo --models gemini-2.5-flash,claude-sonnet-4-20250514`,
+      `  agentaudit audit https://github.com/owner/repo --models opus,sonnet,grok-4,gemini-3.1-pro`,
       `  agentaudit audit https://github.com/owner/repo --format sarif > results.sarif`,
       `  agentaudit audit https://github.com/owner/repo --export`,
     ],
@@ -5565,7 +5583,7 @@ async function main() {
     console.log(`    agentaudit discover --quick`);
     console.log(`    agentaudit scan https://github.com/owner/repo`);
     console.log(`    agentaudit audit https://github.com/owner/repo`);
-    console.log(`    agentaudit audit <url> --models gemini-2.5-flash,claude-sonnet-4-20250514`);
+    console.log(`    agentaudit audit <url> --models opus,sonnet,grok-4,gemini-3.1-pro`);
     console.log(`    agentaudit lookup fastmcp --json`);
     console.log();
     console.log(`  ${c.bold}LEARN MORE${c.reset}`);
@@ -6087,6 +6105,13 @@ async function main() {
       console.log(`  Active:  ${c.yellow}no provider configured${c.reset}`);
     }
     console.log();
+    console.log(`  ${c.dim}Recommended for deep audits (validated on real-world CVEs):${c.reset}`);
+    console.log(`  ${c.dim}  Anthropic  claude-opus-4    — best precision${c.reset}`);
+    console.log(`  ${c.dim}  Anthropic  claude-sonnet-4  — best value${c.reset}`);
+    console.log(`  ${c.dim}  xAI        grok-4           — complementary detection${c.reset}`);
+    console.log(`  ${c.dim}  Google     gemini-3.1-pro   — deepest analysis${c.reset}`);
+    console.log(`  ${c.dim}  For multi-model consensus:  agentaudit audit <url> --models opus,sonnet,grok-4,gemini-3.1-pro${c.reset}`);
+    console.log();
 
     // Step A: Provider selection
     const providerChoices = [

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentaudit",
-  "version": "3.13.4",
+  "version": "3.13.6",
   "description": "Security scanner for AI agent packages — CLI + MCP server",
   "type": "module",
   "bin": {

package/prompts/audit-prompt.md

@@ -158,7 +158,7 @@ For each evidence item from Phase 2, apply the following checks IN ORDER.
 
 | # | Question | If YES → |
 |---|----------|----------|
-| 1 | Is this the package's documented core functionality? (Check Package Profile "Expected Behaviors") | **NOT a finding** (or at most LOW/by_design). See Core-Functionality-Exemption below. |
+| 1 | Is this the package's documented core functionality AND does the implementation properly validate/sanitize its inputs? | Only if BOTH are true → **at most LOW/by_design**. If the feature is core but has implementation flaws (missing input validation, unsanitized paths, no access control) → it IS a finding at normal severity. See Core-Functionality-Exemption below. |
 | 2 | Do I have a specific file:line:code snippet as evidence? | If NO → **DO NOT report**. Speculative findings are never findings. |
 | 3 | Is this a `.env`, `.env.example`, or `process.env`/`os.environ` pattern for self-configuration? | **NOT a finding** (unless the credential is exfiltrated to an external endpoint). |
 | 4 | Can I write a concrete 2-sentence attack scenario? | If NO → **Maximum severity LOW**. |
@@ -238,7 +238,7 @@ A package that integrates multiple APIs requiring multiple credentials is a feat
 - Negation contexts ("never use eval"), install docs (`sudo apt`)
 
 ### ❌ Opt-In Features with Safety Warnings ≠ Default Vulnerabilities
-If a feature must be EXPLICITLY enabled (via env var, config flag, CLI option) AND the naming/docs warn about risks, this is NOT a vulnerability in the default configuration.
+If a feature must be EXPLICITLY enabled (via env var, config flag, CLI option) AND the naming/docs warn about risks, the EXISTENCE of that feature is NOT a vulnerability.
 ```
 ❌ FALSE POSITIVE: MCP server has ENABLE_UNSAFE_SSE_TRANSPORT env var (default: unset/disabled) → NOT Critical (at most LOW/by_design)
 ❌ FALSE POSITIVE: Helm chart has useLegacyRules: false with documented "not recommended for production" → NOT a finding (defaults are safe)
@@ -248,6 +248,8 @@ If a feature must be EXPLICITLY enabled (via env var, config flag, CLI option) A
 ```
 **Key distinction:** "Vulnerable if operator explicitly opts in" (LOW/by_design) vs "Vulnerable by default" (HIGH/CRITICAL). Count the prerequisites — each explicit opt-in step REDUCES severity.
 
+**IMPORTANT:** Opt-in reduces severity for the EXISTENCE of a feature, but NOT for implementation flaws WITHIN the feature. If an opt-in feature has missing input validation, path traversal, or injection vulnerabilities, those are still findings at normal severity when the feature is enabled.
+
 ### ❌ Secure Code Patterns ≠ Injection Vulnerabilities
 These code patterns are SECURE and must NOT be flagged:
 ```
@@ -271,7 +273,18 @@ If you cannot find the EXACT code pattern in the provided source files, do NOT r
 If the pattern is in the Package Profile's "Expected Behaviors" list:
 - It **CANNOT** be MEDIUM or higher severity
 - It is either **NOT a finding** or at most **LOW / by_design**
-- **EXCEPTIONS** (still flag even if expected): Unescaped identifier interpolation, missing parameterization of VALUES, missing operation allowlists
+- **EXCEPTIONS** (still flag at normal severity even if expected behavior):
+  - Unescaped identifier interpolation, missing parameterization of VALUES, missing operation allowlists
+  - **Missing input validation/sanitization on file paths** (path traversal in file-handling tools)
+  - **Missing input validation on tool arguments** that reach dangerous sinks (exec, fs.write, network calls)
+  - **Missing access control on destructive operations** (delete, overwrite, install)
+
+**IMPORTANT — Feature vs Implementation Flaw:**
+A feature can be "expected behavior" while still having implementation flaws. Distinguish:
+- "The server writes files" → expected, by_design
+- "The server writes files to unsanitized user-controlled paths" → vulnerability (PATH_TRAV)
+- "The server installs extensions" → expected, by_design
+- "The server installs extensions from any arbitrary path without validation" → vulnerability (SEC_BYPASS)
 
 ## 3.4 Credential-Config-Normalization (Hard Rule)
 
@@ -290,7 +303,28 @@ If the pattern is in the Package Profile's "Expected Behaviors" list:
 2. Credentials are logged/printed at INFO level or higher in production code paths
 3. Credentials are sent to unexpected external endpoints (exfiltration)
 
-## 3.5 Exploitability Assessment (Mandatory for every candidate)
+## 3.5 MCP Trust Boundary Rule
+
+**In MCP servers, ALL tool arguments from the LLM/client are UNTRUSTED input.**
+
+Even though the LLM is the primary caller, tool arguments may originate from:
+- Prompt injection attacks (malicious content in documents, web pages, emails)
+- Compromised or malicious MCP clients
+- Multi-agent systems where upstream agents are untrusted
+
+Therefore: any tool argument that reaches a dangerous sink (file system, shell, network, database) without validation IS a vulnerability, regardless of whether the operation is "expected behavior."
+
+**Examples:**
+```
+✅ FINDING: file_write(path=request.params.path) where path is not validated → PATH_TRAV
+✅ FINDING: install_extension(path=request.params.path) without path restriction → SEC_BYPASS
+✅ FINDING: execute_query(sql=request.params.query) with string interpolation → SQL injection
+❌ NOT A FINDING: execute_query(sql) where sql is passed as parameterized value
+❌ NOT A FINDING: file_write(path) where path is validated against allowlist/root directory
+```
+
+## 3.6 Exploitability Assessment (Mandatory for every candidate)
+
 
 For each candidate finding, evaluate:
 
@@ -313,7 +347,7 @@ For each candidate finding, evaluate:
 
 **If you cannot describe a concrete 2-sentence attack scenario, the finding is NOT CRITICAL or HIGH.**
 
-## 3.6 Devil's Advocate (Mandatory for HIGH and CRITICAL)
+## 3.7 Devil's Advocate (Mandatory for HIGH and CRITICAL)
 
 Before any finding becomes HIGH or CRITICAL, you MUST argue AGAINST it:
 
@@ -326,7 +360,7 @@ DEVIL'S ADVOCATE:
 
 If the counter-argument is stronger than the finding → demote or exclude.
 
-## 3.7 Reasoning Chain (Mandatory for HIGH and CRITICAL)
+## 3.8 Reasoning Chain (Mandatory for HIGH and CRITICAL)
 
 Every HIGH or CRITICAL finding MUST include this explicit reasoning:
 
@@ -342,7 +376,7 @@ THEREFORE: severity = [X]
 
 If you cannot complete steps 3 or 5, demote to MEDIUM or lower.
 
-## 3.8 Severity Assignment
+## 3.9 Severity Assignment
 
 ### Severity Anchoring
 
@@ -408,7 +442,7 @@ If you cannot complete steps 3 or 5, demote to MEDIUM or lower.
 
 If data collection or exfiltration is gated behind CI environment variables (`process.env.CI`, `GITHUB_ACTIONS`, `JENKINS_URL`, `TRAVIS`, `CIRCLECI`, `GITLAB_CI`), escalate findings within the CI-gated block by one severity level. A legitimate library has no reason to conditionally activate data collection only in CI. Only escalate findings whose code is inside or triggered by the CI-conditional block.
 
-## 3.9 By-Design Classification
+## 3.10 By-Design Classification
 
 A finding is `by_design: true` ONLY when ALL FOUR are true:
 1. **Core purpose**: Pattern is essential to documented purpose (not side-effect)
@@ -436,7 +470,7 @@ If **any** fails → real vulnerability (`by_design: false`).
 - **Development-mode fallbacks** (e.g. fallback JWT secret when env var is not set, localhost-only defaults): Standard in web frameworks. If the fallback only activates in development/missing-config scenarios and production requires explicit configuration → `by_design: true`.
 - **Transparent monetization** (e.g. referral fees, affiliate links, commission systems): If the package EXPLICITLY documents its monetization model in README/SKILL.md and the user can see it before using → `by_design: true`. The finding is still valuable as information but should not count against trust score. Note: UNDISCLOSED affiliate links (hidden in URLs without documentation) are NOT by_design.
 
-## 3.10 Final Triage
+## 3.11 Final Triage
 
 ### Finding Quality Check
 
@@ -649,7 +683,7 @@ Consult these patterns during Phase 2 evidence collection. Remember: a pattern m
 - **Insecure protocols** (`SEC_BYPASS_005`): HTTP for sensitive data.
 - **Overly broad permissions** (`PRIV_ESC_003`): Read all files/env/network when not needed.
 - **Unsafe deserialization (local)** (`DESER_001`): `pickle.loads()`, `yaml.load()` without safe loader on LOCAL data. Remote source → CRITICAL.
-- **Path traversal** (`PATH_TRAV_001`): Unsanitized `../` in paths.
+- **Path traversal** (`PATH_TRAV_001`): Unsanitized user-controlled path reaching filesystem operations. Patterns: `path.resolve(userInput)` + `fs.writeFile`, `fs.readFile(userInput)`, `fs.rm(userInput)`. The `../` characters need NOT be present in the source — the vulnerability is that the user CAN supply them at runtime.
 - **Weak crypto** (`CRYPTO_WEAK_001`): MD5/SHA1 for security, hardcoded IVs. Always report as separate finding.
 - **Capability escalation**: Instructions to "enable dev mode", "unlock capabilities", "bypass restrictions".
 - **Context pollution**: "remember forever", "inject into context", "prepend to every response".
@@ -691,11 +725,12 @@ Consult these patterns during Phase 2 evidence collection. Remember: a pattern m
 
 1. Tool descriptions/schemas — hidden instructions or prompt injection?
 2. Transport config — `npx -y` without version pinning?
-3. File access tools — path sanitization?
+3. **File access tools — path sanitization?** Trace EVERY tool handler that takes a file path, directory path, or URL from `request.params` → check if it reaches `fs.writeFile`, `fs.readFile`, `fs.rm`, `path.resolve`, `path.join` WITHOUT validation against an allowlist or root directory constraint.
 4. Permissions — minimal scope, documented?
 5. Descriptions match code behavior?
-6. Arguments passed to `exec()`/`system()` without sanitization?
+6. **Arguments passed to `exec()`/`system()`/`installExtension()`/dangerous sinks without sanitization?** MCP tool arguments are untrusted. Trace from `request.params.*` → to execution. ANY unsanitized path is PATH_TRAV, ANY unsanitized string in exec is CMD_INJECT.
 7. Error messages — info leaks or injection payloads?
+8. **Unrestricted destructive operations** — delete, overwrite, install operations that take user-controlled targets without access control or scope restriction.
 
 ---
 
@@ -711,6 +746,19 @@ Consult these patterns during Phase 2 evidence collection. Remember: a pattern m
 4. **`bgauryy--octocode-mcp`**: Shell injection via `execAsync()` with shell-string interpolation of `symbolName` in lspReferencesPatterns.ts:317. ✅ HIGH correct.
 5. **`mendez1212--automation-workflows`**: Obfuscated Lua malware payload with luajit dropper. ✅ CRITICAL correct — 10/10 findings valid.
 
+## Real-World True Positives (Confirmed by manual code review)
+
+6. ✅ **`chrome-devtools-mcp`**: "Arbitrary file write due to unsanitized user-controlled paths" rated HIGH. Reality: `path.resolve(filename)` + `fs.writeFile(filePath, data)` at McpContext.ts:840 — writes to any path without validation. Attack: prompt injection causes LLM to call file_write with `../../etc/crontab`. ✅ HIGH correct.
+7. ✅ **`chrome-devtools-mcp`**: "Arbitrary Chrome extension installation via tool parameter" rated HIGH. Reality: `context.installExtension(path)` takes unsanitized path from MCP tool arg. Even though gated behind experimentalExtensionSupport, when enabled there's no path restriction. ✅ HIGH correct (when feature enabled).
+8. ✅ **`terraform-mcp-server`**: "Insecure TLS verification enabled for Terraform Cloud client" rated HIGH. Reality: TLS verification disabled in registry client (registry.go). ✅ HIGH correct.
+9. ✅ **`terraform-mcp-server`**: "Unverified external binary download and execution in CI workflow" rated HIGH. Reality: CI workflow downloads binary without checksum verification. ✅ HIGH correct.
+10. ✅ **`mcp-grafana`**: "Insecure TLS certificate verification bypass enabled by flag" rated MEDIUM. Reality: Flag allows disabling TLS verification for Grafana API calls. ✅ MEDIUM correct.
+
+**Key lesson from these real-world TPs:**
+- A feature being "expected behavior" does NOT exempt it from input validation requirements
+- MCP tool arguments are untrusted — missing validation on paths/URLs reaching fs/exec IS a finding
+- Opt-in features can still have implementation vulnerabilities that should be flagged
+
 ## Incorrect Findings (False Positives — DO NOT repeat)
 
 1. ❌ **`video-transcript`**: "Shell RC File Modification for Persistence" rated CRITICAL. Reality: Adds PATH entry to `.bashrc` — standard installation, not malware. Should be LOW at most.