full_agent_scanLABSCO SUMMARY
A checklist for the failure modes that only exist once a model can act.
Prompt injection, poisoned tool descriptions and excessive agency are risks a conventional application scan will not look for. Having them as callable checks means an agent design gets reviewed while it is being assembled rather than after something happens.
What it is
A security-assessment server built on the OWASP Top 10 for Agentic AI. It reviews an agent setup against the categories that apply to agents rather than to web applications: prompt injection, tool poisoning, excessive agency and data leakage.
What you get
- A full sweep across an agent configuration (full_agent_scan)
- An OWASP-framed assessment of an agentic setup (assess_agentic_security)
- A prompt-injection check (check_prompt_injection)
- A tool-poisoning check over the tool definitions an agent is exposed to (check_tool_poisoning)
- An excessive-agency check on how much the agent is permitted to do unsupervised (check_excessive_agency)
Requirements
Python and pip install owasp_agentic_mcp, or the Smithery installer. The tools themselves need no key; the vendor's paid Pro and Enterprise plans cover the wider MCP suite and EU AI Act tracking rather than these checks.
Setup effort
One command — pip install owasp_agentic_mcp
USE CASES
01
Reviewing an agent before you give it real permissions
Run the full sweep over the configuration and read the OWASP-framed assessment alongside it, so the decision to grant a capability is made against a named category rather than a hunch.
02
Vetting a third-party server before you install it
Point the tool-poisoning check at the tool definitions the server publishes. Descriptions are instructions the model reads, and that is exactly where an unhelpful one hides.
03
Testing whether a prompt can be turned against you
Run the prompt-injection check on the surface where untrusted text reaches the model, which is where a support ticket, a web page or a file comment becomes a command.
04
Deciding how much the agent should be allowed to do alone
The excessive-agency check reads the permissions the agent holds and reports where the reach exceeds the job, which is the argument you need to take one of them away.
5 tools
Listed in project docs. Your client may show a different list.
assess_agentic_securitycheck_prompt_injectioncheck_tool_poisoningcheck_excessive_agencyBEFORE YOU INSTALL
What to know first
Tools you get
5 tools named in the project's documentation.
Requires
Python 3.10+
Where it runs
All 4 documented clients run the server on your machine.
INSTALLATION
Connect your client
Server setup
TERMINAL
pip install owasp_agentic_mcpMCP Clients
Claude Desktop, Cursor, Windsurf
JSON
{
"mcpServers": {
"owasp-agentic": {
"command": "owasp_agentic_mcp"
}
}
}VS Code
VS CODE · JSON
{
"servers": {
"owasp-agentic": {
"command": "owasp_agentic_mcp"
}
}
}
