diff --git a/data/mcp/mcp_hardcoded_secrets.yaml b/data/mcp/mcp_hardcoded_secrets.yaml
new file mode 100644
index 000000000..7d8d5162b
--- /dev/null
+++ b/data/mcp/mcp_hardcoded_secrets.yaml
@@ -0,0 +1,82 @@
+info:
+  id: "mcp_hardcoded_secrets"
+  name: "MCP Hardcoded Secrets Detection"
+  description: "Detect hardcoded credentials, API keys, tokens, and private keys embedded in MCP server source code, which can be exposed if the repository or distribution package is accessible to unauthorized parties."
+  author: "ATR (Agent Threat Rules)"
+  categories:
+    - code
+
+prompt_template: |
+  As a professional AI-agent security analyst, precisely detect hardcoded secrets
+  in MCP server source code. Hardcoded credentials embedded directly in source
+  files — API keys, access tokens, private keys, passwords — can be extracted by
+  anyone who gains read access to the repository, package, or running container.
+  Require concrete evidence of a genuine secret value; only report when the string
+  matches a known credential pattern with sufficient entropy or format.
+
+  ## Vulnerability Definition
+  Hardcoded secrets occur when sensitive authentication material — cloud provider
+  keys (`AKIA…`), API tokens (`sk-…`, `ghp_…`, `xoxb-…`), private keys
+  (`-----BEGIN … PRIVATE KEY-----`), database passwords, JWT secrets — is
+  committed directly in source code rather than loaded from environment variables,
+  a secrets manager, or a properly git-ignored configuration file.
+
+  ## Detection Criteria (require a concrete secret pattern)
+
+  ### 1. Cloud provider credentials
+  **Code Patterns:**
+  - AWS access keys: `AKIA[0-9A-Z]{16}`, AWS secret keys assigned as string
+    literals.
+  - Azure / GCP keys, service account JSON embedded as a string or object.
+  - Cloudflare / DigitalOcean / Linode API tokens in assignments.
+
+  ### 2. SaaS API tokens
+  **Code Patterns:**
+  - OpenAI / Anthropic / other LLM API keys: `sk-[a-zA-Z0-9]{20,}`,
+    `sk-ant-…`.
+  - GitHub tokens: `ghp_[a-zA-Z0-9]{36}`, `gho_…`, `ghs_…`.
+  - Slack tokens: `xox[baprs]-…`.
+  - Stripe keys: `sk_live_…`, `rk_live_…`.
+  - Generic `API_KEY = "…"`, `SECRET_KEY = "…"`, `TOKEN = "…"` assignments
+    with high-entropy values (>20 chars, mixed case + digits).
+
+  ### 3. Private keys and certificates
+  **Code Patterns:**
+  - PEM blocks: `-----BEGIN RSA PRIVATE KEY-----`, `-----BEGIN EC PRIVATE KEY-----`,
+    `-----BEGIN PRIVATE KEY-----`, `-----BEGIN OPENSSH PRIVATE KEY-----`.
+  - PGP private key blocks, certificate bundles with private key material.
+
+  ### 4. Database / service passwords
+  **Code Patterns:**
+  - Connection strings with embedded passwords:
+    `mongodb://user:password@host`, `postgres://user:pass@host`.
+  - Password assignments: `password = "…"`, `DB_PASS = "…"` where the value is
+    a non-placeholder string (not `<password>`, `your_password`, `changeme`,
+    `example`, `xxx`).
+
+  ## Strict Judgment Standards
+  - **Environment variables**: Do not report `os.environ.get('API_KEY')` or
+    `process.env.API_KEY` — these are correct practices.
+  - **Placeholders**: Do not report obvious placeholder values (`your_key_here`,
+    `xxx`, `changeme`, `example`, `<token>`, empty strings).
+  - **Test fixtures**: Do not report secrets in files under test directories
+    (`test/`, `__tests__/`, `*_test.*`, `*.spec.*`) that are clearly fake values.
+  - **Documentation**: Do not report secrets inside comments that serve as
+    usage examples with placeholder-style values.
+  - **Low-entropy values**: Do not report short or dictionary-word strings that
+    do not match any credential format.
+
+  ## Output Requirements
+  Only output when finding concrete hardcoded secrets:
+  - Specific file path and line number
+  - The type of secret (API key, token, private key, password, etc.)
+  - The redacted value (show only enough to identify the type, e.g.
+    `AKIA****XXXX`)
+  - Technical analysis: how the secret could be extracted by an unauthorized party
+  - Impact assessment: what resources or services the exposed credential grants
+    access to
+  - Remediation: move secrets to environment variables or a secrets manager,
+    rotate the exposed credential, add `.env` to `.gitignore`, use
+    pre-commit secret scanners.
+
+  **Strict Requirement: provide the file, line number, and secret type with a redacted value. Remain silent when no concrete evidence exists.**
diff --git a/data/mcp/mcp_insecure_deserialization.yaml b/data/mcp/mcp_insecure_deserialization.yaml
new file mode 100644
index 000000000..20e8dd13e
--- /dev/null
+++ b/data/mcp/mcp_insecure_deserialization.yaml
@@ -0,0 +1,75 @@
+info:
+  id: "mcp_insecure_deserialization"
+  name: "MCP Insecure Deserialization Detection"
+  description: "Detect MCP server tools that deserialize caller-supplied data using unsafe serializers (pickle, marshal, eval-based JSON, yaml.load, unserialize) without validation, enabling remote code execution on the MCP host."
+  author: "ATR (Agent Threat Rules)"
+  categories:
+    - code
+
+prompt_template: |
+  As a professional AI-agent security analyst, precisely detect insecure
+  deserialization vulnerabilities in MCP server tools. The agent (or an upstream
+  prompt-injection) controls tool arguments; if those arguments are deserialized
+  using an unsafe serializer, the MCP host can execute attacker-crafted payloads.
+  Require a concrete argument-to-deserialization-sink path; only report with
+  evidence.
+
+  ## Vulnerability Definition
+  An MCP tool handler takes caller-supplied arguments (strings, base64 blobs, or
+  nested objects) and passes them to a deserializer that can instantiate
+  arbitrary objects or execute code — `pickle.loads`, `marshal.loads`,
+  `yaml.load` (unsafe), `unserialize` (PHP), `ObjectInputStream` (Java),
+  `eval`/`Function()` on JSON-like strings — without type/schema validation.
+  Because tool arguments are model/attacker-influenced, this is argument-driven
+  RCE on the MCP host.
+
+  ## Detection Criteria (require tainted argument -> deserialization sink)
+
+  ### 1. Python unsafe deserializers
+  **Code Patterns:**
+  - `pickle.loads(args.data)` / `pickle.load(open(args.file))` on
+    caller-controlled input.
+  - `marshal.loads(args.data)` / `yaml.load(args.data)` without `Loader=SafeLoader`
+    / `yaml.unsafe_load()`.
+  - `dill.loads()` / `shelve.open()` on attacker-influenced data.
+  - `jsonpickle.decode(args.data)` on caller-controlled input — jsonpickle can
+    instantiate arbitrary classes via `{"py/object": "..."}` directives,
+    carrying the same RCE risk as pickle.
+
+  ### 2. JavaScript / dynamic evaluation as deserialization
+  **Code Patterns:**
+  - `eval(args.expr)` / `new Function(args.code)` used to parse a structured
+    argument.
+  - `vm.runInNewContext(args.script)` / `vm.Script` on caller input.
+  - `JSON.parse` followed by `eval` of a contained field.
+
+  ### 3. Other language-specific unsafe deserializers
+  **Code Patterns:**
+  - PHP `unserialize($args['data'])` without `allowed_classes` restriction.
+  - Java `ObjectInputStream.readObject()` on caller-controlled bytes.
+  - .NET `BinaryFormatter.Deserialize()` / `LosFormatter.Deserialize()` on
+    attacker input.
+  - Ruby `Marshal.load(args.data)` / `YAML.load` (unsafe) on caller input.
+
+  ## Strict Judgment Standards
+  - **Safe serializers**: Do not report `json.loads` / `JSON.parse` /
+    `yaml.safe_load` / `yaml.load(..., Loader=SafeLoader)` — these do not
+    instantiate arbitrary objects.
+  - **Trusted data**: Do not report deserialization of server-generated or
+    statically-defined data with no caller influence.
+  - **Type-restricted**: Do not report PHP `unserialize` with `allowed_classes:
+    false` or an explicit allowlist.
+  - **Test/example**: Do not report sample handlers with hardcoded test payloads.
+
+  ## Output Requirements
+  Only output when finding a concrete deserialization path:
+  - Specific file paths and line numbers for the tainted argument and the sink
+  - The tool parameter name and the exact deserialization call
+  - Technical analysis: how a crafted payload achieves code execution or object
+    instantiation
+  - Impact assessment: RCE scope on the MCP host
+  - Remediation: use safe serializers (`json`, `yaml.safe_load`), avoid
+    `eval`/`pickle`/`unserialize` on untrusted data, enforce schema validation,
+    restrict `allowed_classes`.
+
+  **Strict Requirement: provide the tainted-argument source and the deserialization sink with line numbers. Remain silent when no concrete evidence exists.**
diff --git a/data/mcp/mcp_path_traversal.yaml b/data/mcp/mcp_path_traversal.yaml
new file mode 100644
index 000000000..72a2436b0
--- /dev/null
+++ b/data/mcp/mcp_path_traversal.yaml
@@ -0,0 +1,71 @@
+info:
+  id: "mcp_path_traversal"
+  name: "MCP Tool Path Traversal Detection"
+  description: "Detect MCP server tools that accept file or directory path arguments and pass them to filesystem APIs without canonicalization or directory-boundary checks, enabling arbitrary file read/write outside the intended scope."
+  author: "ATR (Agent Threat Rules)"
+  categories:
+    - code
+
+prompt_template: |
+  As a professional AI-agent security analyst, precisely detect path-traversal
+  vulnerabilities in MCP server tools. The agent (or an upstream prompt-injection)
+  controls tool arguments; if a path argument reaches a filesystem API without
+  canonicalization or boundary checks, the MCP server can read or write files
+  outside the intended directory. Require a concrete argument-to-filesystem-sink
+  path; only report with evidence.
+
+  ## Vulnerability Definition
+  An MCP tool handler takes caller-supplied arguments containing a file path,
+  directory, or filename and passes them to filesystem APIs (`fs.readFile`,
+  `fs.writeFile`, `open()`, `os.path.join`, `Path.Combine`, etc.) without
+  resolving the absolute path and verifying it stays within an allowed root
+  directory. Because tool arguments are model/attacker-influenced, this enables
+  arbitrary file access on the MCP host.
+
+  ## Detection Criteria (require tainted argument -> filesystem sink)
+
+  ### 1. Unsanitized path to filesystem API
+  **Code Patterns:**
+  - `fs.readFile(args.path)` / `fs.readFile(path.join(baseDir, args.filename))`
+    where `args.filename` contains `../` sequences that escape `baseDir`.
+  - `open(args.file)` / `os.path.join(workdir, user_input)` with no subsequent
+    canonicalization check.
+  - `Path.Combine(root, userInput)` without verifying `result.StartsWith(root)`.
+  - Filename argument used in write/delete APIs: `fs.writeFile`, `fs.unlink`,
+    `os.remove`, `shutil.rmtree` without boundary enforcement.
+
+  ### 2. Missing canonicalization
+  **Code Patterns:**
+  - No call to `path.resolve()` / `fs.realpath()` / `os.path.abspath()` /
+    `os.path.realpath()` before the filesystem operation.
+  - Canonicalized path not compared against the allowed root with a prefix check
+    (e.g., `resolved.startsWith(allowedRoot + sep)`).
+  - Symlink-following without restriction, allowing escape via symlinks.
+
+  ### 3. Template / pattern injection into paths
+  **Code Patterns:**
+  - Path built via string interpolation: `` `${baseDir}/${args.name}.json` ``
+    where `args.name` contains `/` or `..` to traverse directories.
+  - Glob or regex patterns derived from user input that can match files outside
+    the intended scope.
+
+  ## Strict Judgment Standards
+  - **Fixed paths**: Do not report hardcoded, constant file paths with no argument
+    interpolation.
+  - **Properly validated paths**: Do not report paths that are canonicalized and
+    boundary-checked before the filesystem operation.
+  - **Schema-restricted input**: Do not report arguments constrained by a strict
+    schema (e.g., enum of allowed filenames, no path separators allowed).
+  - **Test/example**: Do not report sample handlers with placeholder paths.
+
+  ## Output Requirements
+  Only output when finding a concrete path-traversal path:
+  - Specific file paths and line numbers for the tainted argument and the sink
+  - The tool parameter name and the exact filesystem call
+  - Technical analysis: how a crafted path escapes the intended directory
+  - Impact assessment: which files or directories could be accessed
+  - Remediation: canonicalize paths with `resolve`/`realpath`, enforce a directory
+    boundary check, reject path separators in filename arguments, use a safe
+    allowlist.
+
+  **Strict Requirement: provide the tainted-argument source and the filesystem sink with line numbers. Remain silent when no concrete evidence exists.**
diff --git a/data/mcp/mcp_ssrf.yaml b/data/mcp/mcp_ssrf.yaml
new file mode 100644
index 000000000..644ead5ff
--- /dev/null
+++ b/data/mcp/mcp_ssrf.yaml
@@ -0,0 +1,69 @@
+info:
+  id: "mcp_ssrf"
+  name: "MCP Server-Side Request Forgery (SSRF) Detection"
+  description: "Detect MCP server tools that accept URL or hostname arguments from the caller and fetch them without validating against internal/private network ranges, enabling server-side request forgery."
+  author: "ATR (Agent Threat Rules)"
+  categories:
+    - code
+
+prompt_template: |
+  As a professional AI-agent security analyst, precisely detect Server-Side Request
+  Forgery (SSRF) vulnerabilities in MCP server tools. The agent (or an upstream
+  prompt-injection) controls tool arguments; if a URL or hostname argument reaches
+  an HTTP client or socket without validation, the MCP server can be coerced into
+  accessing internal services, cloud metadata endpoints, or other restricted
+  resources. Require a concrete argument-to-network-request path; only report with
+  evidence.
+
+  ## Vulnerability Definition
+  An MCP tool handler takes caller-supplied arguments containing a URL, hostname,
+  or IP address and passes them to a network client (`fetch`, `axios`, `requests`,
+  `urllib`, raw socket, etc.) without checking the destination against internal or
+  private address ranges. Because tool arguments are model/attacker-influenced,
+  this enables SSRF from the MCP host.
+
+  ## Detection Criteria (require tainted argument -> network sink)
+
+  ### 1. Unvalidated URL fetch
+  **Code Patterns:**
+  - `fetch(args.url)` / `axios.get(args.url)` / `requests.get(params["url"])`
+    with no prior allowlist or IP-range check on the URL.
+  - URL constructed by concatenating a user-supplied host/path:
+    `` `https://${args.host}/${args.path}` `` passed directly to a client.
+  - Tool argument used as the `baseURL` / `endpoint` of an API client without
+    validation.
+
+  ### 2. Internal/private address access
+  **Code Patterns:**
+  - No blocklist for `127.0.0.0/8`, `10.0.0.0/8`, `172.16.0.0/12`,
+    `192.168.0.0/16`, `169.254.169.254` (cloud metadata), `::1`, `fc00::/7`.
+  - DNS rebinding exposure: hostname validated but resolution not re-checked
+    after DNS lookup.
+  - Redirect following enabled without re-validating the final destination.
+
+  ### 3. Scheme / protocol confusion
+  **Code Patterns:**
+  - No scheme restriction, allowing `file://`, `gopher://`, `dict://`, `ftp://`
+    schemes to reach local files or internal services.
+  - URL parsing that accepts a user-supplied scheme without an `http`/`https`
+    allowlist.
+
+  ## Strict Judgment Standards
+  - **Fixed URLs**: Do not report hardcoded, constant URLs with no argument
+    interpolation.
+  - **Validated destinations**: Do not report URLs passed through a strict
+    allowlist of domains or a properly implemented SSRF guard (IP-range check +
+    scheme check + redirect-following disabled or re-validated).
+  - **Test/example**: Do not report sample handlers with placeholder URLs.
+
+  ## Output Requirements
+  Only output when finding a concrete SSRF path:
+  - Specific file paths and line numbers for the tainted argument and the sink
+  - The tool parameter name and the exact network call
+  - Technical analysis: how a crafted URL reaches an internal or restricted
+    resource
+  - Impact assessment: what internal endpoints or metadata could be accessed
+  - Remediation: validate destinations against an allowlist, block private/loop
+    back ranges, restrict schemes to http/https, disable or re-validate redirects.
+
+  **Strict Requirement: provide the tainted-argument source and the network sink with line numbers. Remain silent when no concrete evidence exists.**