agentlens-scanner

AgentLens

AgentLens is a deterministic, pre-installation security wrapper for AI Agent skills, tools, and repositories.

It forces targeted code to run through a custom static analysis risk scoring model before allowing it to interact with your execution environment, aggressively identifying common vulnerabilities introduced by AI-generated external dependencies.

Key Features

• Deterministic Scanning: Fully bounded risk scoring engine that prevents unbounded linear score accumulation.
• Categorical Risk Silos: Independent evaluation of code_execution, prompt_injection, filesystem_access, and network_access.
• Policy-Driven Decision Engine: Automatically maps combined signals into actionable ALLOW, WARN, or BLOCK decisions based on YAML configuration.
• Context-Aware Scoring: Intelligently extracts structural signals (like is_framework) to separate high-risk raw execution from safe library runtime internals.
• Confidence Scoring: Validates the strength and ambiguity of risk signals, gracefully downgrading uncertain blocks to warnings.
• LLM Semantic Analysis: Optional second-opinion pass powered by Azure AI Foundry that evaluates the true intent of flagged code snippets, reducing false positives without sacrificing security coverage.
• Behavioral Analysis (NEW): Optional deep-scan layer that detects dynamic imports, runtime code execution (exec/eval), obfuscation patterns (base64+exec), and exfiltration domains. Safely unpacks and analyzes .whl and .tar.gz archives.
• Registry Targets: Scan packages directly with npm:<package>, pypi:<package>, or clawhub:<skill> (downloads are extracted with path-traversal checks on archives).
• JS/TS Guardrails: JavaScript and TypeScript files are scanned for high-risk execution patterns, and npm: packages with JS/TS sources are never treated as clean solely because the Python AST analyzer found nothing.
• Output Formats: Rich CLI formatting (yielding clear human-readable explanations and recommendations), or full Pydantic-validated JSON for programmatic aggregation.

Installation

Install from PyPI:

pip install agentlens-scanner

Install with the optional prompt-injection prefilter:

pip install "agentlens-scanner[injection]"

Install from source:

git clone https://github.com/ellacarmon/AgentLens.git
cd AgentLens
pip install .

Usage

Scan targets

• GitHub: HTTPS URL to a repository (must be github.com).
• Local path: Directory or file on disk.
• npm: npm:<package> (e.g. npm:lodash) — fetches the latest tarball from the public registry.
• PyPI: pypi:<package> (e.g. pypi:requests) — prefers an .tar.gz sdist, otherwise a .whl.
• ClawHub: clawhub:<skill> (e.g. clawhub:calendar-helper) — downloads the published skill ZIP from ClawHub's public API.

Examples:

agentlens scan https://github.com/langchain-ai/langchain
agentlens scan ./local_skill_folder
agentlens scan npm:some-package
agentlens scan pypi:some-project
agentlens scan clawhub:some-skill

Check the installed CLI version:

agentlens --version

To integrate into programmatic pipelines (such as a GitHub action or a pre-flight execution check), use --json:

agentlens scan https://github.com/microsoft/autogen --json

For pipeline safety, AgentLens automatically returns semantic exit codes reflecting the decision engine state:

• 0: ALLOW
• 1: WARN
• 2: BLOCK

To enforce custom strict policies, provide a custom YAML policy template:

agentlens scan ./local_skill_folder --policy custom_policy.yml

GitHub Actions example: scan pull requests

This repository includes an example workflow at .github/workflows/agentlens-pr-scan.yml that scans pull requests on opened, synchronize, reopened, and ready_for_review.

It checks out the PR branch, installs agentlens-scanner, runs a repository scan with JSON output, and uploads the report as a workflow artifact:

- name: Scan repository contents
  id: agentlens
  shell: bash
  run: |
    set +e
    agentlens scan . --json > agentlens-report.json
    exit_code=$?
    echo "exit_code=$exit_code" >> "$GITHUB_OUTPUT"
    exit 0

The example treats BLOCK as a failing check and leaves WARN visible in the workflow summary without failing the job. If you want warnings to block merges too, change the final enforcement step to fail on exit code 1 as well.

GitHub Actions example: scan pinned AI dependencies from requirements files

This repository also includes .github/workflows/agentlens-requirements-ai-scan.yml, which looks for requirements*.txt files in a pull request, extracts pinned Python dependencies, filters them to a curated set of AI-related package names, and scans each pinned version as a pypi: target.

Example requirements.txt entries that this workflow will pick up:

openai==1.68.2
langchain==0.3.21
transformers==4.49.0

The example intentionally only scans exact == pins so the workflow analyzes the same version that would be installed from the file. Unpinned entries such as openai>=1.0 or non-AI packages are ignored.

LLM Semantic Analysis (Azure AI Foundry)

The static analysis engine is fast and deterministic, but can produce false positives — for example, flagging a legitimate subprocess call used for a local math calculation the same way it flags a reverse shell. The semantic analysis layer adds a second-opinion pass that uses an LLM to evaluate the true intent of the flagged code snippet.

When enabled, the hybrid engine runs the static tier first. If the result is WARN or BLOCK and there is at least one code_execution or network_access finding, the LLM is invoked once per scan on a small batch of the strongest such findings (up to three), so the model sees a bit of cross-file context instead of a single line in isolation. A high-confidence ALLOW verdict from the LLM overrides the static decision, updating both the final verdict and the recommendation.

To print which findings were sent to the model (paths, rules, severities), run the CLI with verbose on the top-level command: agentlens -v scan ... --semantic.

Setup

Set the following environment variables:

export AZURE_OPENAI_API_KEY=your-key
export AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
export AZURE_OPENAI_API_VERSION=2024-12-01-preview  # optional, this is the default

Usage

Add --semantic to any scan to enable the hybrid engine:

agentlens scan https://github.com/langchain-ai/langchain --semantic

Override the default deployment name or confidence threshold:

agentlens scan https://github.com/langchain-ai/langchain --semantic --semantic-model gpt-4o --semantic-threshold 0.90

When a semantic override is applied, the output includes a dedicated section:

Semantic Analysis:
  Decision: ALLOW
  Confidence: 0.95
  Explanation: The subprocess call is used for a bounded local computation with no external I/O.
  Flagged Pattern: subprocess.run with shell=True

If the LLM API is unavailable or returns an unparseable response, AgentLens degrades gracefully to the static verdict — the scan never fails due to an LLM outage.

Behavioral Analysis

Behavioral analysis extends AgentLens beyond static AST pattern matching to detect dynamic code execution, runtime module loading, and obfuscation techniques that would otherwise evade traditional scanning.

What It Detects

The behavioral analyzer identifies:

1. Dynamic Imports: Runtime module loading via __import__(), importlib.import_module(), or obfuscated getattr(importlib, ...) patterns
2. Runtime Code Execution: exec(), eval(), and compile() calls, with severity escalation for dynamic (variable) arguments
3. Obfuscation Patterns:
   • Base64-encoded Python code
   • Base64 decode combined with exec/eval (classic obfuscation)
   • Excessive getattr() usage indicating attribute-based obfuscation
4. Suspicious Behavioral Patterns:
   • Network requests to exfiltration domains (pastebin, webhooks, ngrok tunnels, etc.)
   • File writes to suspicious locations (/tmp, home directory)
5. Archive Unpacking: Safely extracts and analyzes .whl and .tar.gz packages with path-traversal protection

When to Use It

Enable behavioral analysis when:

• Scanning PyPI wheel packages or tarballs that may contain compiled/obfuscated code
• Investigating skills that use plugin systems or dynamic loading
• Dealing with packages that have low static risk but suspicious architectural patterns
• You need deeper inspection beyond surface-level AST analysis

Note: Behavioral analysis adds 2-5 seconds per scan and is disabled by default for performance.

Usage

Add --behavioral to any scan:

# Scan a local skill with behavioral analysis
agentlens scan ./suspicious_plugin --behavioral

# Scan a PyPI package (automatically unpacks wheel if needed)
agentlens scan pypi:analytics-helper --behavioral

# Combine with semantic analysis for maximum coverage
agentlens scan ./plugin --behavioral --semantic

When behavioral analysis is enabled, the output includes a dedicated summary:

Behavioral Analysis:
  Findings: 7
  Dynamic Imports: 3
  Runtime Execution: 2
  Obfuscation: 1
  Suspicious Patterns: 1
  Archive Unpacked: Yes

Behavioral Detection Rules

The analyzer uses the following rule IDs (see agentlens/rules/policy.yaml for full details):

• BEH-001: Dynamic import via __import__()
• BEH-002: Dynamic import via importlib.import_module()
• BEH-003: Obfuscated dynamic import via getattr(importlib, ...)
• BEH-004: Runtime code execution via exec()
• BEH-005: Runtime code execution via eval()
• BEH-006: Dynamic code compilation via compile()
• BEH-007: Suspicious exfiltration domain detected
• BEH-008: Base64 decode + exec pattern (obfuscation)
• BEH-009: Suspicious file write location
• BEH-010: Base64-encoded Python code
• BEH-011: Excessive getattr() usage

Risk Scoring Impact

Behavioral findings contribute to the overall risk score through the feature-driven scoring system:

• Dynamic Imports (variable module names): +8.5 to code_execution category
• Runtime Execution (exec/eval with dynamic args): +9.5 to code_execution category
• Obfuscation (base64+exec): +9.0 to code_execution category
• Exfiltration Domains: +8.0 to network_access category

These scores combine with static analysis findings through probabilistic OR aggregation to produce the final risk score.

Safety & Performance

Security Guarantees:

• No Code Execution: The analyzer NEVER executes untrusted code. All analysis is static AST parsing only.
• Zip/Tar Bomb Protection:
  • Maximum extracted size: 500MB
  • Maximum single file: 100MB
  • Maximum compression ratio: 100:1
  • Maximum file count: 10,000
• Path Traversal Prevention: All archive members validated before extraction
• Symlink Protection: Symlinks are detected and skipped entirely
• Special File Blocking: Device files, FIFOs, and other special files are rejected
• Filename Validation: Null bytes, control characters, and oversized names blocked
• Extraction Isolation: Each member's final path is verified to stay within temp directory
• Temporary Isolation: Archives unpacked to isolated temp directories with automatic cleanup
• Timeout Protection: Per-analysis timeout of 5 seconds prevents hanging
• Graceful Degradation: If behavioral analysis fails, the scan continues with static analysis only

Performance:

• Adds 2-5 seconds per scan
• Disabled by default (opt-in with --behavioral)
• Efficient AST parsing with minimal overhead

Releasing To PyPI

This repository includes an automated publish workflow at .github/workflows/publish-pypi.yml.

On every push to main, GitHub Actions will:

• derive a unique package version for that commit from the base major.minor in pyproject.toml
• run the test suite
• build the wheel and source distribution
• validate the artifacts with twine check
• publish to PyPI with Trusted Publishing

The generated version format is:

<major>.<minor>.<github_run_number>

If a workflow run is retried, it publishes a unique post-release for that same commit:

<major>.<minor>.<github_run_number>.post<N>

One-time PyPI setup

Configure a Trusted Publisher in PyPI for this repository:

• Owner: your GitHub user or org
• Repository: ellacarmon/AgentLens
• Workflow name: publish-pypi.yml
• Environment: leave empty unless you later restrict publishing with a GitHub Environment

No long-lived PyPI API token is required when Trusted Publishing is enabled.

Contributing

Please see CONTRIBUTING.md for details on submitting pull requests to the project.

Research

The archetype-based detection in AgentLens (Data Thief and Agent Hijacker) is grounded in:

Malicious Agent Skills in the Wild: A Large-Scale Security Empirical Study
Liu et al., 2026 · arXiv:2602.06547

This paper provides the first large-scale empirical study of malicious AI agent skills, confirming 157 malicious skills across 98,380 samples. Key findings that informed our detection model:

• 84.2% of vulnerabilities reside in natural-language skill documentation (SKILL.md), not executable code.
• The ecosystem splits into two negatively-correlated archetypes: Data Thieves (credential harvest + remote script execution, empirical OR=556 — a statistical Odds Ratio indicating these behaviors are 556x more likely to occur together) and Agent Hijackers (instruction override + autonomy suppression).
• Advanced attacks use shadow features — capabilities present at runtime but absent from public documentation — in 100% of Level 3 (sophisticated) cases.

License

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details.