NVIDIA

NVIDIA/SkillSpector

Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, security risks, prompt injection, data exfiltration, and supply-chain risks in Claude Code, Codex, and MCP skills before you install them.

View on GitHub Website / Demo Back to listing

24.5kstars 2.1kforks 100+contributors Python

README.md

View on GitHub

# SkillSpector

**Security scanner for AI agent skills.** Detect vulnerabilities, malicious patterns, and security risks before installing agent skills.

[![Python 3.12+](https://img.shields.io/badge/python-3.12+-blue.svg)](https://www.python.org/downloads/)
[![License: Apache 2.0](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://www.apache.org/licenses/LICENSE-2.0)

## Overview

AI agent skills (used by Claude Code, Codex CLI, Gemini CLI, etc.) execute with implicit trust and minimal vetting. In the 31,132-skill analyzed subset of the research dataset, **26.1% of skills contain vulnerabilities** and **5.2% show likely malicious intent**.

SkillSpector helps you answer: **"Is this skill safe to install?"**

SkillSpector is part of the [NVIDIA Verified Skills pipeline](https://docs.nvidia.com/skills/), which scans, evaluates, and signs agent skills before publication. Skills that pass are published to the [NVIDIA skills catalog](https://github.com/NVIDIA/skills).

## Documentation

- **[Scan agent skills before installation](https://docs.nvidia.com/skills/scanning-agent-skills)** — Hosted guide: when to scan, how to read a report, and how to gate installs.
- **[Development guide](docs/DEVELOPMENT.md)** — Architecture, package layout, and how to extend the analyzer pipeline.
- **[Analysis resource bounds](docs/ANALYSIS_RESOURCE_BOUNDS.md)** — Fail-closed bundle, parser, nested-artifact, ledger, and finding ceilings.
- **[Pi extension](docs/PI_EXTENSION.md)** — Install SkillSpector as a Pi tool for scanning skills from inside agent sessions.
- **[OpenCode extension](docs/OPENCODE_EXTENSION.md)** — Install SkillSpector as an OpenCode tool and `/skillspector` command for scanning skills from inside agent sessions.

## Features

- **Multi-format input**: Scan Git repos, URLs, zip files, directories, or single files
- **71 vulnerability patterns** across 17 categories: prompt injection, data exfiltration, privilege escalation, supply chain, excessive agency, output handling, system prompt leakage, memory poisoning, tool misuse, rogue agent, anti-refusal, trigger abuse, dangerous code (AST), taint tracking, YARA signatures, MCP least privilege, and MCP tool poisoning
- **Two-stage analysis**: Fast static analysis + optional LLM semantic evaluation
- **Live vulnerability lookups**: SC4 queries [OSV.dev](https://osv.dev) for real-time CVE data with automatic offline fallback
- **Multiple output formats**: Terminal, JSON, Markdown, and SARIF reports
- **Risk scoring**: 0-100 score with severity labels and clear recommendations
- **Baseline / false-positive suppression**: Accept known findings via a glob-rule or fingerprint baseline so re-scans surface only *new* issues ([docs](docs/SUPPRESSION.md))

## Quick Start

### Installation

> **Open-source software notice:** This project will download and install additional third-party open source software projects. Review the license terms of these open source projects before use.

Create and activate a virtual environment first (all `make` targets assume the venv is active). Use **uv** or **pip**; the Makefile uses `uv` if available, otherwise `pip`.

**Quick install with uv (CLI-only):**

```bash
uv tool install git+https://github.com/NVIDIA/skillspector.git
# Update later: uv tool update skillspector
```

If you plan to run `skillspector mcp`, install the MCP extra at install time:

```bash
uv tool install 'skillspector[mcp] @ git+https://github.com/NVIDIA/skillspector.git'
```

**From source:**

```bash
# Clone the repository
git clone https://github.com/NVIDIA/skillspector.git
cd skillspector

# Create and activate virtual environment
uv venv .venv && source .venv/bin/activate
# or: python3 -m venv .venv && source .venv/bin/activate

# Install for production use
make install

# Or install with development dependencies
make install-dev
```

### Docker (no Python required)

Run SkillSpector without installing Python by building it locally from the included [Dockerfile](Dockerfile). The image is based on the Docker Official Python `3.12-slim-bookworm` image.

**Build the image:**

```bash
make docker-build
# or: docker build -t skillspector .
```

**Scan a local directory** by mounting your current directory into `/scan`, the container's working directory:

```bash
docker run --rm -v "$PWD:/scan" skillspector scan ./my-skill/ --no-llm
```

**Scan with LLM analysis** by passing credentials with a local `.env` file:

```bash
cat > .env <<'EOF'
SKILLSPECTOR_PROVIDER=anthropic
ANTHROPIC_API_KEY=sk-ant-...
EOF
```

```bash
docker run --rm \
  -v "$PWD:/scan" \
  --env-file .env \
  skillspector scan ./my-skill/
```

Or pass credentials directly from your shell environment:

```bash
docker run --rm \
  -v "$PWD:/scan" \
  -e SKILLSPECTOR_PROVIDER=anthropic \
  -e ANTHROPIC_API_KEY="$ANTHROPIC_API_KEY" \
  skillspector scan ./my-skill/
```

**Write a report to the host filesystem** by writing to the mounted directory:

```bash
docker run --rm --user "$(id -u):$(id -g)" \
  -v "$PWD:/scan" \
  skillspector scan ./my-skill/ --no-llm --format json --output report.json
```

**Optional alias** for repeated static scans:

```bash
alias skillspector-docker='docker run --rm -v "$PWD:/scan" skillspector'
skillspector-docker scan ./my-skill/ --no-llm
```

### Basic Usage

```bash
# Scan a local skill directory
skillspector scan ./my-skill/

# Scan a single SKILL.md file
skillspector scan ./SKILL.md

# Scan a Git repository
skillspector scan https://github.com/user/my-skill

# Scan a zip file
skillspector scan ./my-skill.zip
```

#### Size limits

SkillSpector enforces two independent caps on remote and archive inputs to bound the impact of oversized downloads and zip bombs:

- **Per-ingest cap**: `INGEST_MAX_BYTES` (100 MiB) — applied to streamed URL downloads, total uncompressed size of zip archives, and post-clone disk usage of Git repos.
- **Zip member cap**: `INGEST_MAX_ZIP_MEMBERS` (10,000) — caps the number of entries in a single zip.

Note that the per-file 1 MB analysis cap (`MAX_FILE_BYTES`) is a separate, downstream limit: it bounds what individual analyzers will read out of an already-ingested directory. The ingest caps above bound how much content can land on disk in the first place. A breach of either ingest cap fails closed with an `IngestLimitExceededError`.

### Output Formats

```bash
# Terminal output (default) - pretty formatted
skillspector scan ./my-skill/

# JSON output - machine readable
skillspector scan ./my-skill/ --format json --output report.json

# Markdown output - for documentation
skillspector scan ./my-skill/ --format markdown --output report.md

# SARIF output - for CI/CD integration and IDE tooling
skillspector scan ./my-skill/ --format sarif --output report.sarif
```

### Recursive Collection Scanning

`--recursive` scans immediate subdirectories containing `SKILL.md` as independent
skills, with a default aggregate limit of **32 skills**. To raise or lower it:

```bash
skillspector scan ./my-skills/ --recursive --max-skills 100 --no-llm --format json
```

`--max-skills` requires a local recursive collection and a positive integer; `0`
does not mean unlimited. Discovery, byte, finding-record, time and report-size
budgets still apply, so increasing the count does not guarantee complete coverage.
When the count limit is reached, the CLI names the limit and suggests adjusting it
or scanning sub-skills separately. Omitted skills remain visible in the aggregate
report and coverage stays incomplete. Use `--fail-on-incomplete` in CI to fail on
incomplete coverage; the default exit policy is unchanged.

### Batch Scanning

Scan entire directories of skills in parallel from `contrib/batch_scan/`:

```bash
python -m contrib.batch_scan.batch_scan ./my-skills/ --no-llm
python -m contrib.batch_scan.batch_scan ./my-skills/ --workers 20 -f json -o report.json
python -m contrib.batch_scan.batch_scan ./tests/fixtures/ -f terminal --workers 20
```

Supports multilingual detection (zh/ja/ko) and terminal/JSON/Markdown output.

For LLM scans with higher concurrency, configure multiple API keys following
[`.env.example`](contrib/batch_scan/.env.example) — the pool improves throughput
and resilience, provided the keys don't share an account-level rate limit.

See the [contrib guide](contrib/batch_scan/docs/) for details.

> **Note on LLM support:** The default configuration targets DeepSeek as the
> cheapest public option. DeepSeek-Chat is
> [expected to sunset](https://api-docs.deepseek.com/), and the contributor
> does not have hardware to test against local models. The batch scanner was
> originally tested with OpenAI-compatible endpoints — DeepSeek's lack of
> structured-output support required manual JSON-parsing patches. If you can
> contribute a more universal backend (Ollama, vLLM, or a different provider),
> PRs are very welcome.

### Comparing MCP Registry snapshots

Start with raw registry payload captures (`registry-before.json` and
`registry-after.json`). Save a scan report as `previous.json`, then compare a
later scan against that generated local report:

```bash
skillspector scan registry-before.json --mcp-registry --format json --output previous.json
skillspector scan registry-after.json --mcp-registry --format json \
  --mcp-registry-compare previous.json --output compared.json
```

The optional `comparison` object lists added and removed server identities,
changed normalized fields with their previous and current values, and an
`unchanged_count`. Identity is the server name and version, so a new version
appears as an addition and the old version as a removal if it is absent from the
new scan. Acquisition source and scan timestamp are excluded from comparison.
`unmodeled_changes` lists same-identity records whose raw-record hash changed
while normalized fields match; these are not counted as unchanged. This can
indicate a change to fields outside the snapshot model, or array reordering in
the raw record. Package and remote ordering alone is not a normalized field change.
Entries in `changed` may also contain changes outside the snapshot model; inspect
the raw record to see those changes.
Compare reports with the same selection scope: a server absent from the current
input is reported as removed, which does not prove it was removed from a registry.

The comparison file must be a local report produced by this registry mode with
valid normalized snapshots. Malformed, oversized or duplicate server identities
are rejected. This option requires `--mcp-registry` and is separate from the
finding-suppression `--baseline`: it never suppresses findings or changes risk
scores, and it does not fetch or execute listed server endpoints.

Registry mode rejects skill-only controls such as `--use-shipped-baseline`,
`--transitive`, `--transitive-depth`, `--transitive-allow-prefix`, and
`--transitive-deny-prefix` with exit code 2, including an explicitly supplied
default depth (`--transitive-depth 1`). Omit these options for registry scans.

### Suppressing False Positives (baseline)

Suppress known/accepted findings so the risk score reflects only un-triaged
issues and re-scans surface only *new* findings. See the
[suppression guide](docs/SUPPRESSION.md) for the full reference.

```bash
# Accept all current findings into a baseline (run once), then commit it.
skillspector baseline ./my-skill/ -o .skillspector-baseline.yaml

# Scan against the baseline — only NEW findings are reported and scored.
skillspector scan ./my-skill/ --baseline .skillspector-baseline.yaml

# Review what was suppressed (still excluded from the score).
skillspector scan ./my-skill/ --baseline .skillspector-baseline.yaml --show-suppressed
```

A baseline can also use drift-tolerant glob rules (by rule id, file path, or
message) — see [`.skillspector-baseline.example.yaml`](.skillspector-baseline.example.yaml).
Exact fingerprint baselines are evidence-bound: changing the scanned source or
SkillSpector version keeps the finding active until it is reviewed again.
When a selected baseline or baseline output is stored inside the skill
directory, SkillSpector excludes that exact file from content analysis so its
suppression text cannot create findings or enter regenerated fingerprints;
sibling files remain in normal scan scope.

### Explicit scan scope

For a local directory containing one `SKILL.md` or `skill.md`, repeat `--exclude` to omit
selected files before content analysis:

```bash
skillspector scan ./my-skill --exclude 'tests/*' --exclude 'fixtures/*.json'
```

Patterns are case-sensitive globs matched against the entire relative POSIX path;
`*` also matches `/`. Quote patterns so your shell does not expand them. This is
an explicit caller option, not an author-controlled ignore file. Patterns matching
either manifest name (`SKILL.md` or `skill.md`) are rejected, as are recursive, multi-skill, transitive, registry, and
non-directory inputs. No-match patterns are still recorded in the report.

Files already outside the scan inventory, such as policy-excluded dependencies,
retain their existing policy handling and are not counted as caller exclusions.
Explicitly excluded inventory files remain in the coverage denominator as entirely uninspected. The
report lists the applied patterns, excluded count and individual skipped paths;
any matched exclusion makes coverage partial and prevents a `SAFE` recommendation.
Use `--fail-on-incomplete` when partial scope should fail CI. These exclusions do
not suppress findings in included files or override existing safety limits.

### LLM Analysis

For the best results, configure an OpenAI-compatible LLM endpoint for
semantic analysis. Pick a provider with `SKILLSPECTOR_PROVIDER`; hosted providers ship bundled default models, while CLI providers fall back to the local runtime's default model unless `SKILLSPECTOR_MODEL` is set. SkillSpector also works against
local OpenAI-compatible servers (Ollama, vLLM, llama.cpp) and managed
inference gateways.

| Provider (`SKILLSPECTOR_PROVIDER`) | Credential env var | Endpoint | Default model |
| ---------- | ---- | ---- | ---- |
| `openai` | `OPENAI_API_KEY` (+ optional `OPENAI_BASE_URL`) | api.openai.com (or any OpenAI-compatible URL) | `gpt-5.4` |
| `anthropic` | `ANTHROPIC_API_KEY` | api.anthropic.com | `claude-opus-4-6` (`meta_analyzer`: `claude-sonnet-4-6`) |
| `anthropic_proxy` | `ANTHROPIC_PROXY_API_KEY` + `ANTHROPIC_PROXY_ENDPOINT_URL` | Any Vertex-style raw-predict proxy | `claude-sonnet-4-6` |
| `bedrock` | `AWS_PROFILE` (optional) + `AWS_REGION` — SigV4 via boto3 | AWS Bedrock Runtime | `us.anthropic.claude-sonnet-4-6` |
| `nv_build` | `NVIDIA_INFERENCE_KEY` | build.nvidia.com | `z-ai/glm-5.3` |
| `gemini` | `GOOGLE_CLOUD_PROJECT` (+ optional `GOOGLE_CLOUD_LOCATION`) via ADC | Google Cloud OpenAI-compatible Gemini endpoint | `gemini-3.8-flash` |
| `ollama` | _(none)_ | `OLLAMA_BASE_URL` (default `http://localhost:11434/v1`) | `llama3.1:8b` |
| `azure_openai` | `AZURE_OPENAI_API_KEY` + `AZURE_OPENAI_ENDPOINT` | Azure OpenAI Service | `gpt-4o` (deployment defaults to the model label) |
| `openai_compatible` | `SKILLSPECTOR_COMPAT_API_KEY` + `SKILLSPECTOR_COMPAT_BASE_URL` | Any OpenAI-compatible endpoint | `llama-3.1-70b-versatile` |
| `claude_cli` | _(none — uses local CLI auth)_ | local `claude` binary | local Claude runtime fallback, or `SKILLSPECTOR_MODEL` |
| `codex_cli` | Disabled | Registered for compatibility; its read-only sandbox permits host-file reads | Use an HTTP API provider or another supported CLI provider |
| `gemini_cli` | _(none — uses local CLI auth)_ | local `gemini` binary | local Gemini runtime fallback, or `SKILLSPECTOR_MODEL` |
| `opencode_cli` | _(none — uses local CLI auth)_ | local `opencode` (verified 1.18.31–1.18.35, 1.19.0–1.19.4) binary | local OpenCode runtime fallback, or `SKILLSPECTOR_MODEL` |

For `openai_compatible` gateways requiring extra HTTP headers, set
`SKILLSPECTOR_COMPAT_HEADERS='{"X-Tenant-ID":"my-tenant","X-API-Key":"gateway-key"}'`.
It must be a JSON object with string keys and values and applies to sync and async requests.
Unset or blank values add no headers; invalid configuration is rejected without exposing header values.
Treat this environment variable as a secret when it contains gateway API keys.

`gpt-6.1-sol`, `claude-opus-5-5`, and `claude-sonnet-5-5` remain opt-in through
`SKILLSPECTOR_MODEL` or per-slot overrides. Promoting them to defaults requires
live scanner acceptance covering benign input, credential disclosure, prompt
injection, refusals, and incomplete responses.
Record findings, completion status, latency, and usage, and compare against
previous defaults where accessible.

For NVIDIA Build's `z-ai/glm-5.3`, SkillSpector requests `high` reasoning effort.
Set `SKILLSPECTOR_REASONING_EFFORT` to override it with `low`, `high`, or `max`.
The bundled 128,000-token context and 32,000-token output budgets are conservative
application limits; they do not claim the hosted endpoint's maximum capacity.

Structured output is requested through LangChain's `with_structured_output`.
The default is client-specific: `ChatOpenAI` uses `json_schema`, while other
clients may use tool calling. Some models reject a forced tool call with
HTTP 400 (`tool_choice: type "tool" and "any" are not supported for this
model`). The `anthropic` and `anthropic_proxy` providers route those models
(`claude-opus-5-5`, `claude-sonnet-5-5`, `claude-fable-5-1`, `claude-mythos-5-1`,
or any registry entry with `structured_output: json_schema`) to the native
JSON-schema response format.
Bedrock has no JSON-schema output for them, so the `bedrock` provider leaves
`toolChoice` at `auto`, asks for the tool call in the prompt, and retries a
prose answer; it recognises the model from the model ID, a geo/global
inference-profile ID, or a foundation-model / inference-profile ARN. An
application-inference-profile ARN hides the model, so add that ARN to the
registry (`SKILLSPECTOR_MODEL_REGISTRY`) with `tool_choice: auto`.
The `openai_compatible` provider honours the same `tool_choice: auto` entry
for endpoints that ignore both `response_format` and a forced `tool_choice`
and answer in prose (for example iFlytek's `spark-x2.5`, which is bundled).
The `anthropic`, `anthropic_proxy`, `openai`, and `openai_compatible` providers
also recognise those Claude models inside gateway model IDs such as
`anthropic/claude-opus-5-5` or `aws/anthropic/bedrock-claude-opus-5-5`. Some
OpenAI-compatible gateways turn a `json_schema` response format for Claude into
a forced tool call, so the `openai` and `openai_compatible` providers bind the
schema as a tool with `tool_choice` left at `auto` and retry a prose answer.
`SKILLSPECTOR_STRUCTURED_OUTPUT_METHOD=json_schema|function_calling`
overrides the method for any provider.
The method precedence is `SKILLSPECTOR_STRUCTURED_OUTPUT_METHOD`, the provider's
method hint, an analyzer preference, then the client default. TP4 prefers
`function_calling` only for provider `openai` and the exact model label
`azure/anthropic/claude-opus-5`, with reasoning/thinking controls unset; dated
or suffixed labels and other providers keep their existing behavior. An explicit
environment override or provider hint still wins when reasoning is configured.

```bash
# Stock OpenAI
export SKILLSPECTOR_PROVIDER=openai
export OPENAI_API_KEY=sk-...
skillspector scan ./my-skill/

# Anthropic
export SKILLSPECTOR_PROVIDER=anthropic
export ANTHROPIC_API_KEY=sk-ant-...
skillspector scan ./my-skill/

# Anthropic via Vertex-style proxy (corporate gateways, GCP Vertex AI)
export SKILLSPECTOR_PROVIDER=anthropic_proxy
export ANTHROPIC_PROXY_ENDPOINT_URL=https://my-gateway.example.com/models/claude-sonnet-4-6:streamRawPredict
export ANTHROPIC_PROXY_API_KEY=your-bearer-token
export SKILLSPECTOR_MODEL=claude-sonnet-4-6
skillspector scan ./my-skill/

# AWS Bedrock (Claude via SigV4)
export SKILLSPECTOR_PROVIDER=bedrock
# Optional: select an AWS named profile. When unset, the standard
# boto3 credential chain (env vars, instance metadata, SSO, etc.) resolves.
# export AWS_PROFILE=my-profile
export AWS_REGION=us-west-2  # default if unset
# Default model: us.anthropic.claude-sonnet-4-6
# Override with any Bedrock model ID, cross-region inference-profile
# ID, or your own application-inference-profile ARN. Claude 5.5 is opt-in:
# us./eu. profiles keep requests in that geography, global. routes worldwide.
# export SKILLSPECTOR_MODEL=us.anthropic.claude-sonnet-5-5
skillspector scan ./my-skill/

# NVIDIA build.nvidia.com
export SKILLSPECTOR_PROVIDER=nv_build
export NVIDIA_INFERENCE_KEY=nvapi-...
skillspector scan ./my-skill/

# Gemini on Google Cloud (Application Default Credentials / Workload Identity)
# Prerequisites:
#   1. Google Cloud project with billing enabled.
#   2. Enable Gemini Enterprise Agent Platform / Vertex AI API: `aiplatform.googleapis.com`.
#   3. IAM permission: grant `roles/aiplatform.user` (or at minimum `aiplatform.endpoints.predict`)
#      to your user account or Kubernetes service account.
#   4. Local authentication: run `gcloud auth application-default login`.
#      Configure a quota project if needed: `gcloud auth application-default set-quota-project PROJECT_ID`.
#   5. Kubernetes / GKE: configure Workload Identity and leave GOOGLE_APPLICATION_CREDENTIALS unset
#      rather than exporting service account keys.
# Note on Data Residency:
#   The default `global` endpoint does not support data-residency requirements. While you can target
#   `us`, `eu`, or regional endpoints (e.g. `us-central1`), endpoint selection alone does not guarantee
#   data residency or in-region processing without appropriate organizational policies. Always verify
#   that your selected model is supported in your target location.
# Optional credentials:
#   GOOGLE_APPLICATION_CREDENTIALS is optional and can reference Workload or Workforce Identity Federation
#   configuration files; exporting long-lived service account keys is discouraged.
export SKILLSPECTOR_PROVIDER=gemini
export GOOGLE_CLOUD_PROJECT=my-project-id
# export GOOGLE_CLOUD_LOCATION=global  # default is global; or us, eu, or specific region (e.g. us-central1)
# Default model: gemini-3.8-flash
# export SKILLSPECTOR_MODEL=gemini-3.7-flash
skillspector scan ./my-skill/

# Local Claude CLI — no API key; uses your existing `claude auth login` session
# Requires: claude CLI installed and authenticated (claude auth login)
export SKILLSPECTOR_PROVIDER=claude_cli
# Uses the local Claude CLI runtime fallback unless SKILLSPECTOR_MODEL is set.
# export SKILLSPECTOR_MODEL=claude-sonnet-4-6
skillspector scan ./my-skill/

# Gemini (via OpenAI compatibility layer)
export SKILLSPECTOR_PROVIDER=openai
export OPENAI_API_KEY="YOUR_GEMINI_API_KEY"
export OPENAI_BASE_URL="https://generativelanguage.googleapis.com/v1beta/openai/"
export SKILLSPECTOR_MODEL=gemini-3.5-flash
skillspector scan ./my-skill/

# Local Ollama — no API key
export SKILLSPECTOR_PROVIDER=ollama
# export OLLAMA_BASE_URL=http://localhost:11434/v1  # shown default
export SKILLSPECTOR_MODEL=llama3.1:8b
skillspector scan ./my-skill/

# Azure OpenAI
export SKILLSPECTOR_PROVIDER=azure_openai
export AZURE_OPENAI_API_KEY=...
export AZURE_OPENAI_ENDPOINT=https://example.openai.azure.com/
export AZURE_OPENAI_DEPLOYMENT=my-deployment
skillspector scan ./my-skill/

# Any other OpenAI-compatible endpoint
export SKILLSPECTOR_PROVIDER=openai_compatible
export SKILLSPECTOR_COMPAT_API_KEY=...
export SKILLSPECTOR_COMPAT_BASE_URL=https://api.groq.com/openai/v1
export SKILLSPECTOR_MODEL=llama-3.1-70b-versatile
skillspector scan ./my-skill/

# Override the provider's default model
export SKILLSPECTOR_MODEL=gpt-5.2
skillspector scan ./my-skill/

# Skip LLM analysis (faster, static analysis only)
skillspector scan ./my-skill/ --no-llm
```

### MCP Server

Run SkillSpector as a [Model Context Protocol](https://modelcontextprotocol.io)
server so local MCP-capable agents (Claude Code, Codex CLI, Gemini CLI)
can call scanning as a tool and **gate skill/MCP installs on the
result** — turning SkillSpector into a runtime guardrail instead of an
out-of-band audit step.

`skillspector mcp` requires `skillspector[mcp]`.

```bash
# Install, or reinstall if you already used the CLI-only path
uv tool install --force 'skillspector[mcp] @ git+https://github.com/NVIDIA/skillspector.git'

# FastMCP stdio transport for local CLI agents
skillspector mcp

# Streamable HTTP transport on a local loopback interface
skillspector mcp --transport http --host 127.0.0.1 --port 8000
```

The stdio transport is the current FastMCP path for local CLI agents, and the
initialize hang reported in issue #199 still applies there.

The server exposes a single tool:

- **`scan_skill(target, use_llm=true, output_format="json")`** — scans a Git
  URL, file URL, `.zip`, `.md` file, or directory and returns a structured
  verdict: `risk_score` (0-100), `severity`, `recommendation`,
  `safe_to_install`, and `findings`. It also reports `llm_used` / `scan_mode`
  so a low score from a static-only scan is never mistaken for a clean full
  scan.

```bash
claude mcp add skillspector -- skillspector mcp
```

> **Security — HTTP transport trust model**
>
> The HTTP transport ships **without authentication**. Any caller that can
> reach the port can invoke `scan_skill`. HTTP bindings are restricted to
> loopback IPs (`127.0.0.1` or `::1`); `localhost` binds to `127.0.0.1` without
> DNS resolution. Wildcard, routable and other hostname bindings are rejected.
>
> - For remote access, put an authenticating reverse proxy (e.g. nginx + mTLS)
>   in front of the loopback listener.
> - Local paths and `file://` URLs are **automatically rejected** over HTTP to
>   prevent unauthenticated callers from reading arbitrary host files. Only
>   remote Git and `.zip` URLs are accepted.
>
> HTTP scans accept only credential-free `https://` targets. Git scans never use
> the server's Git login or Git config. If a proxy or custom CA is needed, set
> `HTTPS_PROXY`, `NO_PROXY` or `GIT_SSL_CAINFO` in the server environment.
> Use stdio or the CLI for private repositories. Git redirects are rejected;
> use the current repository URL after a rename or transfer.

## Vulnerability Patterns

SkillSpector detects **71 vulnerability patterns** across 17 categories:

### Prompt Injection (6 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| P1 | Instruction Override | HIGH | Commands to ignore safety constraints |
| P2 | Hidden Instructions | HIGH | Malicious directives in comments/invisible text |
| P3 | Exfiltration Commands | HIGH | Instructions to transmit context externally |
| P4 | Behavior Manipulation | MEDIUM | Subtle instructions altering agent decisions |
| P5 | Harmful Content | CRITICAL | Instructions that could cause physical harm |
| P9 | Whitespace Padding | MEDIUM | Large whitespace padding hiding instructions below/beside the visible area |

### Anti-Refusal (3 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| AR1 | Refusal Suppression | HIGH | Instructions to never refuse or always comply (e.g. "never refuse", "always comply") |
| AR2 | Disclaimer Suppression | HIGH | Instructions to omit warnings, disclaimers, or ethical commentary (e.g. "no disclaimers", "do not moralize") |
| AR3 | Safety Policy Nullification | HIGH | Jailbreak framing that nullifies guardrails (e.g. "you have no restrictions", "ignore your guidelines", "do anything now") |

### Data Exfiltration (4 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| E1 | External Transmission | MEDIUM | Sending data to external URLs |
| E2 | Env Variable Harvesting | HIGH | Enumerating, copying, or searching environment data to collect secrets |
| E3 | File System Enumeration | MEDIUM | Scanning directories for sensitive files |
| E4 | Context Leakage | HIGH | Transmitting conversation context externally |

### Privilege Escalation (3 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| PE1 | Excessive Permissions | LOW | Requesting access beyond stated functionality |
| PE2 | Sudo/Root Execution | MEDIUM | Invoking elevated system privileges |
| PE3 | Credential Access | HIGH | Reading SSH keys, tokens, passwords |

### Supply Chain (10+ patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| SC1 | Unpinned Dependencies | LOW | No version constraints on packages |
| SC2 | External Script Fetching | HIGH | curl \| bash and remote code execution |
| SC3 | Obfuscated Code | HIGH | Base64/hex encoded execution |
| SC4 | Known Vulnerable Dependencies | HIGH | Dependencies with known CVEs (live OSV.dev lookup) |
| SC5 | Abandoned Dependencies | MEDIUM | Unmaintained packages without security updates |
| SC6 | Typosquatting | HIGH | Package names similar to popular packages |
| SC8 | Shipped Python Bytecode | HIGH | `__pycache__` / `.pyc` present (discovery skips; malicious bytecode bypass) |
| SC9 | Concealed Executable Artifact | HIGH | Executable nested in a document container or hidden/disguised artifact; root-level visible source helpers in manifested `.agents/skills/<name>/` and `.claude/skills/<name>/` folders are exempted but remain local-only and are marked as not LLM-reviewed |
| SC10 | Dependency Source Redirection | HIGH | Package-manager source added, replaced, or unresolved |

### Excessive Agency (5 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| EA1 | Unrestricted Tool Access | HIGH | Unfettered tool access without constraints |
| EA2 | Autonomous Decision Making | HIGH | High-impact decisions without human-in-the-loop |
| EA3 | Scope Creep | MEDIUM | Capabilities extending beyond stated purpose |
| EA4 | Unbounded Resource Access | MEDIUM | No rate limits or quotas on resource consumption |
| EA5 | External Model or Provider Selection | MEDIUM/HIGH | Model/provider pins or coding-CLI shell-outs that can switch billing accounts |

### Output Handling (3 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| OH1 | Unvalidated Output Injection | HIGH | Model output used without sanitization |
| OH2 | Cross-Context Output | MEDIUM | Output flows across trust boundaries without validation |
| OH3 | Unbounded Output | MEDIUM | No limits on output size or generation rate |

### System Prompt Leakage (3 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| P6 | Direct Leakage | HIGH | Instructions that expose system prompts or internal rules |
| P7 | Indirect Extraction | MEDIUM | Extraction via rephrasing, translation, or side-channels |
| P8 | Tool-Based Exfiltration | HIGH | System prompts exfiltrated via file writes or network requests |

### Memory Poisoning (3 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| MP1 | Persistent Context Injection | HIGH | Content designed to persist across interactions |
| MP2 | Context Window Stuffing | MEDIUM | Filler content displacing safety constraints |
| MP3 | Memory Manipulation | HIGH | Tampering with agent memory or stored state |

### Tool Misuse (3 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| TM1 | Tool Parameter Abuse | HIGH | Crafted parameters for unintended behavior (shell=True, --force) |
| TM2 | Chaining Abuse | HIGH | Tool chains that bypass individual safety checks |
| TM3 | Unsafe Defaults | MEDIUM | Overly permissive defaults (disabled TLS, no auth) |

### Rogue Agent (2 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| RA1 | Self-Modification | CRITICAL | Modifying own code or configuration at runtime |
| RA2 | Session Persistence | HIGH | Unauthorized persistence via cron jobs or startup scripts |

### Trigger Abuse (3 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| TR1 | Overly Broad Trigger | MEDIUM | Trigger patterns matching common words |
| TR2 | Shadow Command Trigger | HIGH | Triggers that shadow built-in commands or other skills |
| TR3 | Keyword Baiting Trigger | MEDIUM | Generic triggers designed to maximize activation |

### Behavioral AST (9 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| AST1 | exec() Call | CRITICAL | Direct exec() enabling arbitrary code execution |
| AST2 | eval() Call | HIGH | Direct eval() evaluating arbitrary expressions |
| AST3 | Dynamic Import | HIGH | \_\_import\_\_() loading arbitrary modules at runtime |
| AST4 | subprocess Call | HIGH | External command execution via subprocess |
| AST5 | os.system / exec-family | HIGH | Shell commands via os module |
| AST6 | compile() Call | MEDIUM | Code object creation from strings |
| AST7 | Dynamic getattr() | MEDIUM | Arbitrary attribute access with non-literal names |
| AST8 | Dangerous Execution Chain | CRITICAL | exec/eval combined with dynamic source (network, encoded data) |
| AST9 | Reflective getattr() Sink | HIGH | Reflective exec via `getattr(os,'system')` / `getattr(builtins,'exec')` that evades AST1/AST5 |

### Taint Tracking (5 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| TT1 | Direct Taint Flow | HIGH | Data flows directly from a source to a sink without sanitization |
| TT2 | Variable-Mediated Taint Flow | MEDIUM | Data flows from source to sink through intermediate variables |
| TT3 | Credential Exfiltration Chain | CRITICAL | Credentials (env vars, secrets) flow to network output sinks |
| TT4 | File Read to Network Exfiltration | HIGH | File contents flow to network output sinks |
| TT5 | External Input to Code Execution | CRITICAL | Network or user input flows to exec/eval/subprocess sinks |

### YARA Signatures (4 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| YR1 | Malware Match | CRITICAL | YARA rule match for known malware signatures |
| YR2 | Webshell Match | CRITICAL | YARA rule match for webshell patterns |
| YR3 | Cryptominer Match | HIGH | YARA rule match for crypto mining indicators |
| YR4 | Hack Tool / Exploit Match | HIGH | YARA rule match for hack tools or exploit code |

### MCP Least Privilege (4 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| LP1 | Underdeclared Capability | HIGH | Code uses capabilities not listed in declared permissions |
| LP2 | Wildcard Permission | MEDIUM | Permission list contains wildcards (\*, all, full, any) |
| LP3 | Missing Permission Declaration | MEDIUM | No permissions field but code has detectable capabilities |
| LP4 | Overdeclared Permission | LOW | Permission declared but no corresponding code capability found |

### MCP Tool Poisoning (4 patterns)

| ID | Pattern | Severity | Description |
|----|---------|----------|-------------|
| TP1 | Hidden Instructions | HIGH | Hidden directives in metadata (HTML comments, zero-width chars, base64, data URIs) |
| TP2 | Unicode Deception | HIGH | Homoglyphs, RTL overrides, mixed-script identifiers in tool metadata |
| TP3 | Parameter Description Injection | MEDIUM | Injection patterns in parameter definitions (overrides, system tokens, malicious defaults) |
| TP4 | Description-Behavior Mismatch | MEDIUM | Declared tool description does not match actual code behavior (LLM-powered) |

All detected patterns are listed in the tables above.

## Risk Scoring

### Score Calculation

- **CRITICAL issues**: +50 points
- **HIGH issues**: +25 points
- **MEDIUM issues**: +10 points
- **LOW issues**: +5 points
- **Executable scripts**: 1.3x multiplier

### Severity Levels

| Score | Severity | Recommendation |
|-------|----------|----------------|
| 0-20 | LOW | SAFE |
| 21-50 | MEDIUM | CAUTION |
| 51-80 | HIGH | DO NOT INSTALL |
| 81-100 | CRITICAL | DO NOT INSTALL |

## Example Output

### Terminal Output

```
 SkillSpector Security Report  v2.0.0

Skill: suspicious-skill
Source: ./suspicious-skill/
Scanned: 2026-01-29 10:30:00 UTC

Risk Assessment
 Metric          Value
 Score           78/100
 Severity        HIGH
 Recommendation  DO NOT INSTALL

Components (3)
 File              Type      Lines  Executable
 SKILL.md          markdown    142  No
 scripts/sync.py   python       87  Yes
 requirements.txt  text          3  No

Issues (2)

HIGH: Env Variable Harvesting (E2)
    Location: scripts/sync.py:23
    Finding: for key, val in os.environ.items():...
    Confidence: 94%
    Explanation: This code collects environment variables containing
    API keys and secrets, then sends them to an external server.

HIGH: External Transmission (E1)
    Location: scripts/sync.py:45
    Finding: requests.post("https://api.skill.io/env"...
    Confidence: 89%
    Explanation: Data is being sent to an external server. Combined
    with env harvesting above, this indicates credential exfiltration.
```

## Configuration

### Environment Variables

| Variable | Description | Required |
|----------|-------------|----------|
| `SKILLSPECTOR_PROVIDER` | Active LLM provider: `openai`, `anthropic`, `anthropic_proxy`, `bedrock`, `nv_build`, `gemini`, `ollama`, `azure_openai`, `openai_compatible`, `claude_cli`, `gemini_cli`, or `opencode_cli`. Hosted providers use bundled `model_registry.yaml` defaults; CLI providers fall back to the local runtime's default model unless `SKILLSPECTOR_MODEL` is set. Defaults to `nv_build`. | Optional |
| `GOOGLE_CLOUD_PROJECT` | Google Cloud project ID for the `gemini` provider. Authenticates via Google Cloud Application Default Credentials (ADC) or GKE Workload Identity. | Required for LLM analysis when `SKILLSPECTOR_PROVIDER=gemini` |
| `GOOGLE_CLOUD_LOCATION` | Google Cloud location for the `gemini` provider endpoint (e.g. `global`, `us`, `eu`, `us-central1`). Defaults to `global`. | Optional (used when `SKILLSPECTOR_PROVIDER=gemini`) |
| `GOOGLE_APPLICATION_CREDENTIALS` | Optional path to ADC credential/config file (e.g. Workload or Workforce Identity Federation config; exported service account keys are discouraged). For local development, use `gcloud auth application-default login`; for GKE, use Workload Identity. | Optional (used when `SKILLSPECTOR_PROVIDER=gemini`) |
| `NVIDIA_INFERENCE_KEY` | Credential for the `nv_build` provider (build.nvidia.com). | Required for LLM analysis when `SKILLSPECTOR_PROVIDER=nv_build` |
| `OPENAI_API_KEY` | Credential for the OpenAI provider (`SKILLSPECTOR_PROVIDER=openai`). Also serves as the tier-2 fallback in the credential waterfall when the active provider returns no credentials; the fallback then uses the `openai` default model (`gpt-5.4`) for any slot without a model override. Claude Opus or Sonnet 5.5 gateway IDs are not supported through the fallback, because it keeps the active provider's structured-output method; select `SKILLSPECTOR_PROVIDER=openai` for them. | Required for LLM analysis when `SKILLSPECTOR_PROVIDER=openai` |
| `OPENAI_BASE_URL` | Override the OpenAI endpoint (e.g. point at Ollama). | Optional |
| `SKILLSPECTOR_REASONING_EFFORT` | Optional provider- and model-dependent reasoning-effort setting. Non-empty values are trimmed and passed through unchanged. When unset or blank, SkillSpector sends `high` for `nv_build` with `z-ai/glm-5.3`; other provider/model combinations keep their endpoint defaults. The value applies to every model slot, so the `anthropic` `meta_analyzer` default receives the same effort as the analyzers. `gpt-6.1-sol` accepts only `low`, `medium`, `high`, `xhigh`, or `max`; any other value stops a scan before analysis (exit code 2). | Optional |
| `SKILLSPECTOR_OUTPUT_LANGUAGE` | Short, single-line language label (letters, numbers, spaces, `_`, or `-`; maximum 64 characters) for human-readable LLM finding text such as messages, explanations, and remediation. Rule IDs, severity values, paths, code, and other machine-readable values remain unchanged. Unset, blank, or invalid values preserve the default output language. | Optional |
| `SKILLSPECTOR_TEMPERATURE` | Optional sampling temperature from `0` to `1` for hosted providers. Unset or blank preserves the provider default. Lower values can reduce run-to-run variation but do not guarantee identical output. `gpt-6.1-sol` rejects every explicit temperature, including `1`. Claude Opus 4.7 and later, Sonnet 5, Fable 5, Mythos 5 and Haiku 5 models (for example `claude-opus-5-5`, `claude-fable-5-1`, `claude-haiku-5-5`) accept only `1.0` on the native Anthropic endpoint ([API compatibility rule](https://platform.claude.com/docs/en/api/messages/create)); gateways, proxies, and Bedrock still reject every explicit value for these models. Unsupported settings on active LLM slots stop a scan before analysis (exit code 2). Static scans (`--no-llm`) skip provider initialization and control validation. A Bedrock application-inference-profile ARN hides the model name, so declare `sampling: rejected` for it in the model registry. | Optional |
| `SKILLSPECTOR_SEED` | Optional integer sampling seed for OpenAI-compatible and Azure OpenAI providers. Other hosted providers and CLI providers do not receive it. Provider support remains model-dependent; `gpt-6.1-sol` receives it, but OpenAI treats `seed` as best-effort, so output can still differ between runs. | Optional |
| `SKILLSPECTOR_COMPACT_PROMPTS` | Opt-in compact line numbering in LLM prompts: numbered lines render as `L1:`, `L2:` instead of zero-padded `L01:`, `L02:`. Accepted truthy values are `1`, `true`, and `yes` (case-insensitive; surrounding whitespace is trimmed). Unset or any other value keeps the default zero-padded format. | Optional |
| `ANTHROPIC_API_KEY` | Credential for the Anthropic provider (`SKILLSPECTOR_PROVIDER=anthropic`). | Required for LLM analysis when `SKILLSPECTOR_PROVIDER=anthropic` |
| `ANTHROPIC_BASE_URL` | Override the native Anthropic endpoint (default: `https://api.anthropic.com`). | Optional |
| `ANTHROPIC_AUTH_SCHEME` | Set to `bearer` to send `ANTHROPIC_API_KEY` as `Authorization: Bearer` instead of `x-api-key`, for gateways exposing the Messages API. | Optional |
| `ANTHROPIC_PROXY_ENDPOINT_URL` | Full endpoint URL for the Anthropic proxy provider (Vertex-style raw-predict). | Required when `SKILLSPECTOR_PROVIDER=anthropic_proxy` |
| `ANTHROPIC_PROXY_API_KEY` | Bearer token for the Anthropic proxy provider. | Required when `SKILLSPECTOR_PROVIDER=anthropic_proxy` |
| `ANTHROPIC_PROXY_API_VERSION` | `anthropic_version` value sent in the request body (default: `vertex-2023-10-16`). | Optional |
| `AWS_PROFILE` | Named AWS profile for the Bedrock provider — authenticates via SigV4 through boto3. When unset, the standard boto3 credential chain (env vars, instance metadata, SSO, etc.) resolves. | Optional (used when `SKILLSPECTOR_PROVIDER=bedrock`) |
| `AWS_REGION` | AWS region for the Bedrock Runtime endpoint. Defaults to `us-west-2`. | Optional (used when `SKILLSPECTOR_PROVIDER=bedrock`) |
| `OLLAMA_BASE_URL` | Ollama OpenAI-compatible endpoint. Defaults to `http://localhost:11434/v1`. | Optional (used when `SKILLSPECTOR_PROVIDER=ollama`) |
| `AZURE_OPENAI_API_KEY` | API key for the Azure OpenAI provider. | Required when `SKILLSPECTOR_PROVIDER=azure_openai` |
| `AZURE_OPENAI_ENDPOINT` | Azure resource endpoint for the Azure OpenAI provider. | Required when `SKILLSPECTOR_PROVIDER=azure_openai` |
| `AZURE_OPENAI_DEPLOYMENT` | Azure deployment name. Defaults to the selected model label. | Optional |
| `AZURE_OPENAI_API_VERSION` | Azure OpenAI API version. Defaults to `2024-06-01`. | Optional |
| `SKILLSPECTOR_COMPAT_API_KEY` | API key for a generic OpenAI-compatible provider. | Required when `SKILLSPECTOR_PROVIDER=openai_compatible` |
| `SKILLSPECTOR_COMPAT_BASE_URL` | Base URL for a generic OpenAI-compatible provider. | Required when `SKILLSPECTOR_PROVIDER=openai_compatible` |
| `SKILLSPECTOR_MODEL` | Override the active provider model. For hosted providers, this replaces the bundled default from the LLM Analysis table. For CLI providers, this is forwarded as `--model` instead of using the local runtime fallback. | Optional |
| `SKILLSPECTOR_MODEL_<SLOT>` | Override the model for one analyzer slot (for example `SKILLSPECTOR_MODEL_META_ANALYZER`). It takes precedence over `SKILLSPECTOR_MODEL`, which in turn replaces provider slot defaults such as the `anthropic` `meta_analyzer` default. See [model provenance](docs/INFERENCE_USAGE.md#model-provenance) for the full precedence. | Optional |
| `SKILLSPECTOR_MODEL_REGISTRY` | Override the bundled per-provider YAML registry (`src/skillspector/providers/<provider>/model_registry.yaml`) with a custom path. | Optional |
| `SKILLSPECTOR_LOG_LEVEL` | Log level: `DEBUG`, `INFO`, `WARNING`, `ERROR` (default: `WARNING`). | Optional |
| `SKILLSPECTOR_MAX_WORKFLOW_SECONDS` | Aggregate scan deadline in seconds; defaults to `1080` (18 minutes). Accepts a positive finite number. Invalid values warn and retain the default. Set before starting the process; caller-supplied budgets and external tool timeouts still apply. | Optional |
| `SKILLSPECTOR_TP4_MAX_BATCHES` | Maximum description-behavior review batches per TP4 analyzer execution; defaults to `256`. Accepts an integer from `1` to `4096`. Unset or blank uses the default; invalid values warn and retain it. Set before starting the process; file, byte, finding, concurrency, and workflow-time limits still apply. | Optional |
| `SKILLSPECTOR_MAX_DEPENDENCY_SOURCE_ANALYSIS_SECONDS` | Ceiling for the dependency-source analysis pass, in seconds. Defaults to `5.0`, which is the historical value. Raise it on slow or heavily loaded machines so the same unchanged tree does not come back partially inspected, which would set `safe_to_install` to false and record `runtime_limit`. The remaining aggregate workflow time (`SKILLSPECTOR_MAX_WORKFLOW_SECONDS`, 1080 seconds by default) still caps the effective value; values above that remaining time have no effect. Invalid, zero, negative, infinite, and NaN values keep the 5-second default. The setting is resolved when the module is imported, so a new process is required after changing it. See [analysis resource bounds](docs/ANALYSIS_RESOURCE_BOUNDS.md#configuring-the-dependency-source-deadline). | Optional |

> **Disabled provider:** `codex_cli` remains registered but cannot run LLM analysis because its read-only sandbox allows host-file reads. Existing users should select an HTTP API provider or another supported CLI provider.

> **CLI providers** (`claude_cli`, `gemini_cli`, `opencode_cli`): No API key is needed. Authentication is managed entirely by the agent CLI's own login session. SkillSpector never reads or forwards API keys when these providers are active. The subprocess is run with capabilities restricted, and untrusted skill content is delivered only via stdin.
>
> `opencode_cli` currently fails closed unless the installed OpenCode version is one of the individually verified releases (`1.18.31`, `1.18.32`, `1.18.33`, `1.18.34`, `1.18.35`, `1.19.0`, `1.19.1`, `1.19.2`, `1.19.3`, `1.19.4`), whose configuration precedence and deny-all semantics are verified by this release; every other version, including prereleases, still fails closed.

### CLI Options

```bash
skillspector scan --help

Options:
  -f, --format [terminal|json|markdown|sarif]  Output format [default: terminal]
  -o, --output PATH                            Output file path
  --no-llm                                     Skip LLM analysis (static only)
  --yara-rules-dir PATH                        Extra YARA rules directory
  -b, --baseline PATH                          Suppress findings listed in a baseline
  --show-suppressed                            List baseline-suppressed findings
  -V, --verbose                                Show detailed progress
  --help                                       Show this message and exit

# Generate a baseline of all current findings (see docs/SUPPRESSION.md)
skillspector baseline <path> [-o FILE] [--no-llm] [--reason TEXT]
```

Report files and new baselines use private permissions (0600). Existing baselines
retain their access metadata when safely replaced; updates requiring in-place
writes are refused. All file output uses atomic replacement. The parent directory
must be writable; symlinked
parents and non-regular destinations such as devices are refused. On macOS,
ancestor directories must also be readable. File output requires POSIX no-follow
directory operations and is unavailable on Windows. Unsupported file output fails
before analysis. Omit `--output` for scan/batch stdout, or use
`skillspector baseline <path> -o -` for JSON on stdout, then redirect only to a
trusted destination. Run Docker with your user ID as above so you can read its files.

## Integrating SkillSpector

SkillSpector is built to be driven by other tools (CI pipelines, install gates, editor integrations). Its exit code and JSON output are a stable contract.

### Exit codes

`skillspector scan` exits with:

| Code | Meaning |
|------|---------|
| `0` | Scan completed, `risk_score` ≤ 50 (recommendation `SAFE` or `CAUTION`), and no enabled strict gate fired |
| `1` | Scan completed and either `risk_score` > 50, `--fail-on-findings` found an active finding, `--fail-on-incomplete` found partial/incomplete analysis, or `--min-coverage` found coverage below its threshold |
| `2` | Error (bad input, unreadable source, internal failure) |

> By default, the exit code collapses `SAFE` and `CAUTION` into `0`. Use `--fail-on-findings` to gate on any active finding, `--fail-on-incomplete` to gate on incomplete coverage, `--min-coverage PERCENT` to gate on a coverage floor, or read the JSON `recommendation` field for custom policy.

### Machine-readable output

`--format json` produces a JSON report; with no `--output`/`-o` it is written to stdout:

```bash
skillspector scan ./my-skill/ --format json
```

The top-level shape is (this example shows a full LLM-backed scan; with `--no-llm`, `metadata.llm_requested` is `false`):

```json
{
  "skill": { "name": "...", "source": "...", "scanned_at": "<ISO 8601>" },
  "risk_assessment": { "score": 0, "severity": "LOW", "recommendation": "SAFE" },
  "components": [ { "path": "...", "type": "...", "lines": 0, "executable": false, "size_bytes": 0 } ],
  "issues": [ { "id": "...", "category": "...", "severity": "...", "confidence": 0.0, "location": { "file": "...", "start_line": 0 }, "surface": "code" } ],
  "metadata": {
    "has_executable_scripts": false,
    "skillspector_version": "...",
    "llm_requested": true,
    "llm_available": true,
    "inference_usage": [
      {
        "node": "semantic_security_discovery",
        "request_kind": "structured_output",
        "provider": "nv_inference",
        "model": "azure/anthropic/claude-opus-4-6",
        "model_source": "provider_response",
        "usage_source": "provider_response",
        "prompt_tokens": 1000,
        "completion_tokens": 100,
        "cached_tokens": 400,
        "cache_write_tokens": 50,
        "total_tokens": 1100
      }
    ]
  }
}
```

- `risk_assessment.severity` ∈ `LOW | MEDIUM | HIGH | CRITICAL`.
- `risk_assessment.recommendation` ∈ `SAFE | CAUTION | DO_NOT_INSTALL`, mapped from severity: `LOW → SAFE`, `MEDIUM → CAUTION`, `HIGH`/`CRITICAL → DO_NOT_INSTALL`.
- `issues[].surface` and `suppressed[].surface` are advisory, heuristic labels for where a finding landed:
  `code | instructions | docs | tests | comments | config`. JSON expands compacted groups into one `issues[]` entry
  per occurrence, each with its own `surface`; SARIF emits `properties.surface` on each result. The label is
  path/line-based and the scanned author can influence it, so do not use it as the sole suppression or security signal.
- `metadata.llm_error` appears only when LLM analysis was requested but unavailable.
- Archive-member components retain their virtual `path` and include `outer_path`, `nested_path`,
  and container provenance. Virtual paths identify in-memory members, not on-disk files; see
  [JSON component locations](docs/NESTED_ARTIFACT_INSPECTION.md#json-component-locations).
- AE1 findings use **Incomplete referenced artifact analysis**. Their source
  location identifies the reference; `evidence` identifies the affected target,
  analyzer reasons, and available bounds. Review the target's completeness
  ledger when `reasons_truncated` is true. See
  [referenced-artifact diagnostics and Perl help text](docs/ANALYSIS_RESOURCE_BOUNDS.md#diagnosing-incomplete-referenced-artifacts)
  for interpretation and corrective actions.
- `metadata.inference_usage` contains one sanitized record per LLM response when the
  provider exposes token counters. It is an empty list when usage is unavailable;
  SkillSpector never estimates missing tokens. Prompt totals are inclusive of cache
  reads and writes so downstream pricing can separate those partitions safely.
  `model_source` distinguishes an independently identified provider model from
  the exact requested model used when response identity is absent or ambiguous.
  SkillSpector does not currently send Anthropic prompt-cache controls, so its
  scan requests cannot select the separate 5-minute or 1-hour cache-write tiers;
  TTL-specific response fields are normalized defensively into the aggregate
  cache-write counter.
- See [Inference usage telemetry](docs/INFERENCE_USAGE.md) for the complete
  provenance, cache-accounting, privacy, fail-closed ingestion, and downstream
  pricing contract.
- The full per-issue shape is defined by `Finding.to_dict()` in [models.py](src/skillspector/models.py); rely on the fields above and treat any additional fields as best-effort.

For CI/IDE tooling, `--format sarif` emits SARIF 2.1.0. Each result carries its
security severity; rule descriptors carry a score only when all of their results
share that severity. Mixed-severity rules remain unannotated, so consumers that
read only rule-level severity cannot distinguish their result severities.
Recursive scans chain each percent-encoded skill-directory URI through an
absolute scan-root file URI. External dependency locations retain their source
provenance instead of being rebased into the local skill directory.

### Recommended gate mapping

When using SkillSpector as an install gate, map the recommendation to an action:

| `recommendation` | Suggested action |
|------------------|------------------|
| `SAFE` | allow |
| `CAUTION` | prompt / warn the user |
| `DO_NOT_INSTALL` | block |

SkillSpector computes the score band and recommendation; how strict the gate is (e.g. whether `CAUTION` blocks in CI) is a policy decision for the integrating tool.

## Development

### Setup

All `make` targets assume a virtual environment is already created and activated. The Makefile uses **uv** if available, else **pip**.

```bash
# Clone, create venv, activate, install dev dependencies
git clone https://github.com/NVIDIA/skillspector.git
cd skillspector
uv venv .venv && source .venv/bin/activate
# or: python3 -m venv .venv && source .venv/bin/activate
make install-dev

# Run tests
make test

# Run tests with coverage
make test-cov

# Run linting
make lint

# Format code
make format
```

## How It Works

SkillSpector uses a two-stage detection pipeline:

### Stage 1: Static Analysis
- Fast regex-based pattern matching across 11 static analyzers
- AST-based behavioral analysis detecting dangerous calls (exec, eval, subprocess, etc.)
- Live vulnerability lookups via OSV.dev for known CVEs in dependencies
- Scans all analyzer-eligible files in the skill
- High recall (catches most issues)
- Moderate precision (some false positives)

OpenSSF Model Signing / DSSE bundles remain in content analysis, regardless of
filename. Static and LLM analyzers inspect decoded JSON payloads, readable DER
certificate/signature content, and all unknown wrapper fields. Strictly parsed
binary fields are projected rather than treated as executable base64; canonical
bytes remain available to YARA and fingerprints. Unsupported encodings (including
environments without the optional `cryptography` dependency) retain raw fields
and produce an explicit incomplete-analysis warning. Findings in decoded content
point to the original bundle and label their projection in evidence. None of this
verifies signatures, certificate chains, transparency logs, or signer identity.

### Stage 2: LLM Semantic Analysis (Optional)
- Evaluates context and intent
- Confirmed findings may gain an explanation and higher confidence, never lower
- Every deterministic finding stays in the report whether the model confirms it, disputes it, or does not address it
- Findings the model reviews but does not confirm (disputed, low-confidence, or unaddressed) are tagged `llm-unconfirmed` in JSON and SARIF output
- Findings whose review fails are kept without that tag; `evidence.llm_review_outcome` distinguishes `confirmed`, `disagreed`, `low-confidence`, `missing`, and `failed` in JSON and SARIF output
- Results depend on the configured model and prompt context; Stage 2 does not replace
  deterministic findings or guarantee a particular precision rate

The LLM prompt includes anti-jailbreak protections to prevent malicious skills from manipulating the analysis.

## Live Vulnerability Lookups (SC4)

SC4 uses the [OSV.dev](https://osv.dev) API to check dependencies against the full Open Source Vulnerabilities database — covering tens of thousands of advisories across PyPI and npm.

- **No API key required** — OSV.dev is free and unauthenticated.
- **Batch queries** — all dependencies are checked in a single HTTP call.
- **Automatic fallback** — if OSV.dev is unreachable (air-gapped/offline), a small built-in fallback list is used.
- **Caching** — results are cached in-memory for 1 hour to avoid redundant API calls during a session.
- **Pipenv** — reads `Pipfile` package tables and the direct/transitive packages in
  `Pipfile.lock`, including development and custom categories. Exact lock versions
  can resolve unpinned manifest dependencies. Malformed entries and unsupported
  VCS/local sources are reported as incomplete dependency checks.

The tool requires outbound HTTPS access to `api.osv.dev` for live vulnerability data. When that is not available, findings are limited to the static fallback list.

## Locked npm dependency contents (opt-in)

Use `--inspect-dependencies` to statically inspect exact public npm tarballs from
`package-lock.json` or `npm-shrinkwrap.json` v2/v3, including recorded transitive,
development, and optional dependencies:

```bash
skillspector scan ./my-skill --inspect-dependencies --no-llm --fail-on-incomplete
```

Artifacts must match their locked SHA-512 integrity hashes. A disposable Linux
worker inspects supported source files and decoded npm scripts without installing
packages or executing their code. Dependency findings include package coordinates,
an artifact digest, and source locations. Missing artifacts, unsupported sources
or formats, unavailable isolation, and exhausted budgets remain incomplete;
`--fail-on-incomplete` prevents them from passing the scan. Reports also show when
inspection was not requested.

This first version supports the public npm registry and Linux x86-64/AArch64
seccomp isolation. Dependency analysis stays local and uses existing static
data-exfiltration, privilege-escalation, and supply-chain patterns. See
[supported content, resource limits, and coverage boundaries](docs/DEPENDENCY_CONTENT_INSPECTION.md).

## Trust model and data egress

SkillSpector is defense-in-depth, not a sandbox. Know what it does and does not do before relying on it:

- **It never executes the scanned skill.** All analysis is static (regex, Python AST, YARA) plus optional LLM evaluation of file *contents* — the skill's code is never run.
- **LLM analysis sends analyzer-eligible file contents to the configured provider.** When LLM analysis is enabled (the default), file contents are sent to the active `SKILLSPECTOR_PROVIDER` endpoint. DSSE payloads and readable signing metadata are included. Use `--no-llm` to keep contents local (static analysis only).
- **SC4 sends dependency names to OSV.dev.** The supply-chain check queries [OSV.dev](https://osv.dev) with the package names and versions the skill declares, to look up known CVEs. This is fundamental to the check and runs even with `--no-llm`. It sends dependency coordinates (not file contents), requires no API key, and falls back to a bundled list when OSV.dev is unreachable.
- **It does not sandbox the host.** SkillSpector flags risky patterns *before* you install a skill; it does not contain or isolate a skill you choose to install anyway.

## Limitations

- **Non-English content**: May miss patterns in other languages
- **Image-based attacks**: Cannot analyze text in images
- **Encrypted/binary code**: Cannot analyze compiled or encrypted content
- **Runtime behavior**: Static analysis only, no dynamic execution
- **Offline SC4**: Without network access to `api.osv.dev`, SC4 uses a small static fallback list

## Research Background

Based on research from "Agent Skills in the Wild: An Empirical Study of Security Vulnerabilities at Scale" (Liu et al., 2026):

- **Dataset**: 42,447 skills from major marketplaces; 31,132 were analyzed for the following rates
- **Vulnerable**: 26.1% of the analyzed subset contain at least one vulnerability
- **High-severity**: 5.2% of the analyzed subset show likely malicious intent
- **Key finding**: Skills with executable scripts are 2.12x more likely to be vulnerable

## Python API Integration

```python
from skillspector import graph

# Invoke the LangGraph workflow
result = graph.invoke({
    "input_path": "/path/to/skill",
    "output_format": "json",   # terminal, json, markdown, or sarif
    "use_llm": True,           # False for static-only analysis
})

# Access results
print(f"Risk Score: {result['risk_score']}/100")
print(f"Severity: {result['risk_severity']}")
print(f"Recommendation: {result['risk_recommendation']}")

for finding in result["filtered_findings"]:
    print(f"[{finding['severity']}] {finding['rule_id']}: {finding['message']}")
```

## License

Apache License 2.0 - see [LICENSE](LICENSE) for details.

## Contributing

Contributions are welcome! Please read our contributing guidelines and submit pull requests.

## Support

- **Issues**: [GitHub Issues](https://github.com/NVIDIA/skillspector/issues)

Repository info

Default branch: main
Latest version: v2.13.4
Created: 21/03/2026
Last activity: 06/11/2026
License: Apache-2.0
Open issues: 168
Watchers: 83

Topics

agent-security agent-skills agentic-ai ai-security claude-code mcp prompt-injection security-scanner security-tools security-workflow supply-chain-security

Contributors

github-actions[bot]

643 contributions

yashrajp22

288 contributions

rng1995

258 contributions

keshprad

67 contributions

mohgupta-ship-it

63 contributions