About AgentGuard

AgentGuard is an AI Agent Security Scanner built for developers and teams who deploy AI agents in production. AI agents are powerful — they can browse the web, execute code, call APIs, and manage infrastructure — but that power comes with real security risks.

AgentGuard scans your agents in seconds. Paste a configuration, enter a URL, and get a detailed security report with findings ranked by severity and actionable remediation steps.

Why Agent Security Matters

Traditional web application security tools weren't designed for AI agents. Agents introduce new attack vectors — prompt injection, tool abuse, secret leakage through LLM responses, SSRF through tool calls — that require specialized scanning. A single misconfigured agent can expose API keys, leak customer data, or give attackers a foothold into your internal network.

AgentGuard was built specifically for this problem. Every scan runs 12 specialized security modules designed around the unique threat model of AI agents.

Open Source & Local-First

AgentGuard's scanner core is 100% open-source under the Apache-2.0 license. We believe security tools should be auditable, transparent, and under your control. You can inspect every line of scanning logic, verify there are no hidden data collection mechanisms, and contribute improvements back to the community.

Fully Offline Scanning

Install the CLI and scan entirely on your own machine. No data ever leaves your environment — configs, secrets, and scan results stay local.

No Telemetry or Phone-Home

The scanner collects zero telemetry. There are no analytics hooks, no usage reporting, and no required accounts to run local scans.

Audit the Code

Every scanner module is open for review. We encourage security researchers and developers to audit the codebase, report vulnerabilities, and submit improvements.

Security Audit Invitation

We welcome independent security audits of the scanner. If you find an issue, please report it via GitHub Issues or email us directly.

View on GitHub
pip install agentguard-scanner[cli]

Learn more about our open-source packages, CI/CD integration, and how to contribute →

12 Security Modules

Secret Leak Detection

Scans agent configurations and responses for exposed API keys, passwords, tokens, and other credentials. Catches secrets before they reach production or end up in logs.

Prompt Injection Protection

Tests whether your agent can be manipulated into bypassing its system instructions using canary token verification. Embeds unique tokens in 15 injection payloads across 4 categories — direct injection, indirect injection, goal hijacking, and system prompt leakage — and confirms actual exploitation by checking if the agent reproduces the token.

Network Attack Prevention (SSRF)

Checks if your agent can be tricked into making requests to internal networks, cloud metadata endpoints, or other restricted resources. Prevents server-side request forgery through agent tool calls.

Tool Safety Verification

Audits the tools your agent has access to — shell execution, file system, database queries, API calls — and tests whether those tools can be exploited to run unintended commands or access unauthorized data.

Configuration Audit

Examines your agent's configuration files and environment for exposed sensitive settings, overly permissive access controls, debug modes left enabled, and other misconfigurations.

Authentication Security

Comprehensive JWT and authentication testing across 8 vulnerability categories. Tests for missing auth on 14 protected endpoints, expired JWT acceptance, 'none' algorithm bypass, RS256/HS256 algorithm confusion, signature validation failures, session fixation, horizontal privilege escalation (IDOR), and vertical privilege escalation via forged admin claims. Uses hand-crafted JWT tokens and common weak secret dictionary testing to detect real-world authentication flaws.

MCP Security Audit

For agents using the Model Context Protocol, this module tests tool server connections for unauthorized access, excessive permissions, and data exposure through MCP tool definitions.

Endpoint Exposure Scan

Discovers publicly accessible admin panels, debug endpoints, documentation pages, and internal tools that should not be exposed. Finds the attack surface you didn't know existed.

Data Exfiltration Detection

Detects attempts to steal data through encoded URLs, known exfiltration services (webhook.site, requestbin, ngrok), and invisible tracking pixels embedded in agent responses. Catches both direct data theft and covert exfiltration channels.

Typosquatting Detection

Compares package dependencies against 200+ known-good npm and PyPI packages using edit distance analysis. Catches typosquatting attacks where malicious packages impersonate popular ones with subtle misspellings.

Excessive Agency Detection

Tests whether your agent executes destructive actions without asking for confirmation, invokes unauthorized tools, or expands beyond its intended operational scope. Catches agents that blindly comply with dangerous requests instead of requiring human approval.

Canary Token Detection

Plants realistic-looking fake credentials (API keys, database URLs, AWS keys) in the agent's context and checks whether the agent replays them on demand or leaks them into unrelated responses. Detects both direct secret replay and cross-context credential leakage.

How It Works

  1. 1

    Create an account

    Sign up for free — no credit card required. Free accounts get 5 scans per day.

  2. 2

    Submit your agent for scanning

    Paste your agent's configuration (JSON, YAML, or TOML) or enter the URL where it's deployed. AgentGuard supports any agent framework.

  3. 3

    Review your security report

    Each finding includes a severity level (critical, high, medium, low, info), a description of the vulnerability, and step-by-step remediation guidance. Your overall risk score helps you prioritize what to fix first.

Pricing

AgentGuard offers two tiers plus enterprise options:

  • Free — 5 scans per day, all 12 security modules, grade + 3 findings, 30-day history
  • Pro ($29/month) — Unlimited scans, all findings, unlimited history, exports, API access + CI/CD integration, scheduled scans, MD + Skill analysis

Ready to secure your agents?

Start scanning for free — no credit card required.

Get Started Free