feat(static): flag unpinned dependencies across manifest formats

docs/architecture/analyzers/llm-analyzer.md

@@ -109,8 +109,9 @@ analyzer = LLMAnalyzer(model="gemini-2.0-flash-exp", api_key=key)
 # Using LiteLLM format
 analyzer = LLMAnalyzer(model="gemini/gemini-2.0-flash-exp", api_key=key)
 
-# Vertex AI
-analyzer = LLMAnalyzer(model="vertex_ai/gemini-1.5-pro")  # uses GOOGLE_APPLICATION_CREDENTIALS
+# Vertex AI -- uses GOOGLE_APPLICATION_CREDENTIALS if set, otherwise falls
+# back to ambient Application Default Credentials (e.g. Workload Identity)
+analyzer = LLMAnalyzer(model="vertex_ai/gemini-1.5-pro")
 ```
 
 ### Azure OpenAI

docs/architecture/analyzers/static-analyzer.md

@@ -18,7 +18,8 @@ flowchart TD
     A["Manifest validation"] --> B["Instruction body scanning"]
     B --> C["Script/code scanning"]
     C --> D["Consistency checks"]
-    D --> E["Referenced file scanning"]
+    D --> D2["Dependency pinning checks"]
+    D2 --> E["Referenced file scanning"]
     E --> F["Binary file checks"]
     F --> G["Hidden file checks"]
     G --> H["File inventory analysis"]
@@ -41,6 +42,7 @@ Each pass targets a different aspect of the skill package:
 | Instruction body | `_scan_instruction_body()` | SKILL.md content against signature rules |
 | Script scanning | `_scan_scripts()` | Python/bash/other scripts against signatures |
 | Consistency | `_check_consistency()` | Mismatch between manifest claims and actual behavior |
+| Dependency pinning | `_check_dependency_pinning()` | Unpinned dependencies in `requirements*.txt`, `pyproject.toml`, `setup.cfg`, `setup.py`, `Pipfile`, and manifest metadata |
 | Referenced files | `_scan_referenced_files()` | Files mentioned in SKILL.md instructions |
 | Binary files | `_check_binary_files()` | Extension/magic mismatch, archive detection, unknown binaries |
 | Hidden files | `_check_hidden_files()` | Dotfiles, `__pycache__`, policy-allowed exceptions |
@@ -97,6 +99,7 @@ The pack manifest registers all rule sources and metadata for the core detection
 - Hardcoded credentials and secrets
 - Archive/binary risks
 - Tool mismatch and manifest consistency
+- Supply-chain risk from unpinned dependencies
 - Hidden file and dotfile risks
 - Document-embedded threats (PDF, Office macros)
 - Unicode homoglyph attacks

docs/architecture/threat-taxonomy.md

@@ -70,7 +70,7 @@ Skill Scanner currently uses a subset of those codes for agent-skill risk catego
 | Code Execution | `AITech-9.1` | `AISubtech-9.1.1` | Unsafe execution primitives |
 | Obfuscation | `AITech-9.2` | `AISubtech-9.2.1` | Detection-evasion obfuscation patterns |
 | ASCII Smuggling | `AITech-9.2` | `AISubtech-9.2.1` | Unicode Tag Block (U+E0000–U+E007F) used to hide prompt-injection payloads inside skill files; invisible in editors but decoded by LLMs |
-| Supply Chain Attack | `AITech-9.3` | `AISubtech-9.3.1` | Malicious package/tool injection |
+| Supply Chain Attack | `AITech-9.3` | `AISubtech-9.3.1` | Malicious package/tool injection; unpinned dependency versions |
 | Unauthorized Tool Use | `AITech-12.1` | `AISubtech-12.1.3` | Unsafe/undeclared tool execution |
 | Tool Poisoning | `AITech-12.1` | `AISubtech-12.1.2` | Tampering with tool behavior/data |
 | Tool Shadowing | `AITech-12.1` | `AISubtech-12.1.4` | Malicious lookalike/replacement tools |

docs/features/index.md

@@ -297,7 +297,7 @@ The LLM and Meta analyzers work with multiple LLM providers. Select a provider v
 
 ```bash
 pip install cisco-ai-skill-scanner[bedrock]   # AWS Bedrock (IAM credentials)
-pip install cisco-ai-skill-scanner[vertex]    # Google Vertex AI
+pip install cisco-ai-skill-scanner[vertex]    # Google Vertex AI (Workload Identity or service account)
 pip install cisco-ai-skill-scanner[azure]     # Azure OpenAI (managed identity)
 pip install cisco-ai-skill-scanner[all]       # All cloud providers
 ```

docs/reference/configuration-reference.md

@@ -23,7 +23,7 @@ Primary settings for the LLM semantic analyzer.
 
 | Variable | Description | Example |
 |---|---|---|
-| `SKILL_SCANNER_LLM_API_KEY` | Primary API key for LLM analyzer and meta fallback. **(required)** | `sk-ant-...` |
+| `SKILL_SCANNER_LLM_API_KEY` | Primary API key for LLM analyzer and meta fallback. Required for API-key-based providers; not required for Bedrock (IAM), Ollama (local), or Vertex AI (ambient Application Default Credentials). | `sk-ant-...` |
 | `SKILL_SCANNER_LLM_MODEL` | Primary model identifier for semantic analysis. | `anthropic/claude-sonnet-4-20250514` |
 | `SKILL_SCANNER_LLM_PROVIDER` | Optional provider override, including OpenAI-compatible custom endpoint routing. | `openai` |
 | `SKILL_SCANNER_LLM_BASE_URL` | Optional custom endpoint base URL for provider routing. | `https://api.openai.com/v1` |
@@ -58,7 +58,7 @@ Credentials for Vertex AI and Google AI Studio.
 
 | Variable | Description | Example |
 |---|---|---|
-| `GOOGLE_APPLICATION_CREDENTIALS` | Path to GCP service account credentials. | `/path/to/sa-key.json` |
+| `GOOGLE_APPLICATION_CREDENTIALS` | Path to GCP service account credentials. Optional -- if unset, `vertex_ai/*` models fall back to ambient Application Default Credentials (e.g. a GCE/Cloud Run attached service account or Workload Identity), same as Bedrock's IAM role fallback. | `/path/to/sa-key.json` |
 | `GEMINI_API_KEY` | Google AI Studio key; auto-set from `SKILL_SCANNER_LLM_API_KEY` when using Gemini via LiteLLM. | `(auto-set from LLM_API_KEY)` |
 
 ## VirusTotal
@@ -115,7 +115,7 @@ Paths, allowlists, and other advanced settings.
 | `ENABLE_LLM_ANALYZER` | `skill_scanner/config/config.py` |
 | `ENABLE_STATIC_ANALYZER` | `skill_scanner/config/config.py` |
 | `GEMINI_API_KEY` | `skill_scanner/core/analyzers/llm_provider_config.py` |
-| `GOOGLE_APPLICATION_CREDENTIALS` | `.env.example`, `skill_scanner/core/analyzers/llm_provider_config.py` |
+| `GOOGLE_APPLICATION_CREDENTIALS` | `.env.example` |
 | `SKILL_SCANNER_ALLOWED_ROOTS` | `skill_scanner/api/router.py` |
 | `SKILL_SCANNER_LLM_API_KEY` | `.env.example`, `skill_scanner/cli/cli.py`, `skill_scanner/config/config.py`, `skill_scanner/core/analyzer_factory.py`, `skill_scanner/core/analyzers/behavioral_analyzer.py`, `skill_scanner/core/analyzers/llm_analyzer.py`, `skill_scanner/core/analyzers/llm_provider_config.py`, `skill_scanner/core/analyzers/meta_analyzer.py` |
 | `SKILL_SCANNER_LLM_API_VERSION` | `.env.example`, `skill_scanner/cli/cli.py`, `skill_scanner/core/analyzer_factory.py`, `skill_scanner/core/analyzers/meta_analyzer.py` |

docs/reference/dependencies-and-llm-providers.md

@@ -108,7 +108,8 @@ For OpenAI and OpenAI-compatible custom endpoints, `SKILL_SCANNER_LLM_USER` can
 | OpenAI-compatible custom endpoint | API key + endpoint | `SKILL_SCANNER_LLM_API_KEY`, `SKILL_SCANNER_LLM_PROVIDER=openai`, `SKILL_SCANNER_LLM_BASE_URL` |
 | AWS Bedrock (API key) | API key | `SKILL_SCANNER_LLM_API_KEY` |
 | AWS Bedrock (IAM) | AWS credentials | `AWS_REGION`, `AWS_PROFILE` (optional: `AWS_SESSION_TOKEN`) |
-| Google Vertex AI | Service account | `GOOGLE_APPLICATION_CREDENTIALS` |
+| Google Vertex AI (service account) | Service account key file | `GOOGLE_APPLICATION_CREDENTIALS` |
+| Google Vertex AI (ambient ADC) | Workload Identity / attached service account | none -- falls back automatically, like Bedrock IAM |
 | Google AI Studio | API key | `SKILL_SCANNER_LLM_API_KEY` (auto-sets `GEMINI_API_KEY`) |
 | Azure OpenAI | API key + endpoint | `SKILL_SCANNER_LLM_API_KEY`, `SKILL_SCANNER_LLM_BASE_URL`, `SKILL_SCANNER_LLM_API_VERSION` |
 | Ollama | None | — |

pyproject.toml

@@ -76,13 +76,18 @@ dependencies = [
     # Floor pin: avoid compromised PyPI lines in 1.80.16 / 1.82.3 / 1.82.8 (BerriAI/litellm#24512)
     # SSRF fix in 1.83.7 (GHSA-xqmj-j6mv-4862), and CVE-2026-49468 fix in 1.84.0.
     "litellm>=1.84.0,<2",
+    # Floor pin: PYSEC-2026-2132 fixed in 8.3.3. Transitive via litellm/magika/uvicorn/typer.
+    "click>=8.3.3",
 ]
 
 [project.optional-dependencies]
 # Google AI Studio / Gemini SDKs
 google = [
     "google-genai>=1.60,<2",
     "google-generativeai>=0.8,<1",
+    # Floor pins for transitive CVEs pulled in via google-generativeai -> google-api-python-client / google-auth
+    "httplib2>=0.32.0",  # PYSEC-2026-3444
+    "pyasn1>=0.6.4",  # PYSEC-2026-3455, PYSEC-2026-3456, PYSEC-2026-3457
 ]
 # AWS Bedrock support (requires boto3 for IAM credentials)
 bedrock = [
@@ -91,6 +96,9 @@ bedrock = [
 # Google Vertex AI support
 vertex = [
     "google-cloud-aiplatform>=1.130,<2",
+    # Floor pins for transitive CVEs pulled in via google-cloud-aiplatform -> google-api-core / google-auth
+    "httplib2>=0.32.0",  # PYSEC-2026-3444
+    "pyasn1>=0.6.4",  # PYSEC-2026-3455, PYSEC-2026-3456, PYSEC-2026-3457
 ]
 # Azure OpenAI support (for managed identity auth)
 azure = [
@@ -107,6 +115,8 @@ all = [
     "boto3>=1.40,<2",
     "google-cloud-aiplatform>=1.130,<2",
     "azure-identity>=1.20,<2",
+    "httplib2>=0.32.0",  # PYSEC-2026-3444
+    "pyasn1>=0.6.4",  # PYSEC-2026-3455, PYSEC-2026-3456, PYSEC-2026-3457
 ]
 
 [dependency-groups]

scripts/generate_reference_docs.py

@@ -392,7 +392,11 @@ def add(var: str, source: str) -> None:
 
 def _describe_env_var(var: str) -> str:
     descriptions = {
-        "SKILL_SCANNER_LLM_API_KEY": "Primary API key for LLM analyzer and meta fallback.",
+        "SKILL_SCANNER_LLM_API_KEY": (
+            "Primary API key for LLM analyzer and meta fallback. Required for "
+            "API-key-based providers; not required for Bedrock (IAM), Ollama "
+            "(local), or Vertex AI (ambient Application Default Credentials)."
+        ),
         "SKILL_SCANNER_LLM_MODEL": "Primary model identifier for semantic analysis.",
         "SKILL_SCANNER_LLM_PROVIDER": "Optional provider override, including OpenAI-compatible custom endpoint routing.",
         "SKILL_SCANNER_LLM_BASE_URL": "Optional custom endpoint base URL for provider routing.",
@@ -410,7 +414,12 @@ def _describe_env_var(var: str) -> str:
         "AWS_REGION": "AWS region for Bedrock-backed flows.",
         "AWS_PROFILE": "AWS credential profile for Bedrock IAM auth.",
         "AWS_SESSION_TOKEN": "Optional AWS session token.",
-        "GOOGLE_APPLICATION_CREDENTIALS": "Path to GCP service account credentials.",
+        "GOOGLE_APPLICATION_CREDENTIALS": (
+            "Path to GCP service account credentials. Optional -- if unset, "
+            "`vertex_ai/*` models fall back to ambient Application Default "
+            "Credentials (e.g. a GCE/Cloud Run attached service account or "
+            "Workload Identity), same as Bedrock's IAM role fallback."
+        ),
         "SKILL_SCANNER_ALLOWED_ROOTS": "Colon-delimited API path allowlist for server-side path access.",
         "SKILL_SCANNER_TAXONOMY_PATH": "Path to a custom Cisco AI taxonomy YAML file (overridden by `--taxonomy`).",
         "SKILL_SCANNER_THREAT_MAPPING_PATH": "Path to a custom threat mapping YAML file (overridden by `--threat-mapping`).",
@@ -518,7 +527,7 @@ def _describe_env_var(var: str) -> str:
     "SKILL_SCANNER_THREAT_MAPPING_PATH": "/path/to/threats.yaml",
 }
 
-_ENV_VAR_REQUIRED: set[str] = {"SKILL_SCANNER_LLM_API_KEY"}
+_ENV_VAR_REQUIRED: set[str] = set()
 
 
 def _render_configuration_reference() -> str:

skill_scanner/core/analyzers/llm_provider_config.py

@@ -173,7 +173,12 @@ def _resolve_api_key(self, api_key: str | None) -> str | None:
         Uses SKILL_SCANNER_LLM_API_KEY consistently for all providers.
 
         Special cases:
-        - Vertex AI: Uses GOOGLE_APPLICATION_CREDENTIALS (service account)
+        - Vertex AI: Always returns ``None`` -- LiteLLM/google-auth read
+          GOOGLE_APPLICATION_CREDENTIALS directly from the environment when
+          set, or fall back to ambient Application Default Credentials
+          (like Bedrock's IAM role) -- e.g. a GCE/Cloud Run attached
+          service account or Workload Identity, with no key file on disk
+          at all.
         - Ollama: No API key needed (local)
         - Azure: Falls back to Entra ID (``az login``) when no API key is set
         """
@@ -182,7 +187,7 @@ def _resolve_api_key(self, api_key: str | None) -> str | None:
 
         # Special cases with different auth mechanisms
         if self.is_vertex:
-            return os.getenv("GOOGLE_APPLICATION_CREDENTIALS")
+            return None
         elif self.is_ollama:
             return None
 
@@ -269,7 +274,7 @@ def _normalize_gemini_model_name(self, model: str) -> str:
 
     def validate(self) -> None:
         """Validate that configuration is complete."""
-        if not self.is_bedrock and not self.is_ollama and not self.api_key:
+        if not self.is_bedrock and not self.is_ollama and not self.is_vertex and not self.api_key:
             if self.is_azure:
                 raise ValueError(
                     f"No API key or Entra ID credentials found for Azure model {self.model}. "