Agent Immune
Harden agent workflows with prompt-injection detection, semantic memory controls, output scanning, and prompt hardening via an MCP security layer.
Overview
io.github.denial-web/agent-immune is a Ship-phase MCP server that adds prompt-injection detection, output scanning, semantic memory safeguards, and prompt hardening for AI agent deployments.
What is this MCP server?
- Prompt injection detection for untrusted user or tool-origin content
- Semantic memory features positioned for safer long-horizon agent context
- Output scanning to catch risky exfiltration or policy violations before display
- Prompt hardening utilities to reduce jailbreak and instruction-smuggling success
- PyPI package agent-immune v0.2.2 with stdio MCP transport
- PyPI package agent-immune version 0.2.2
- Four advertised capability areas: injection detection, semantic memory, output scanning, prompt hardening
What problem does it solve?
Agents that read untrusted text can be steered by injection attacks and leak harmful outputs, which solo builders often discover only after shipping.
Who is it for?
Indie developers shipping MCP-connected agents that consume user-supplied or web-sourced content and need lightweight guardrails in the same stack.
Skip if: Teams needing formal SOC2/compliance suites only, or projects with no agent surface and no untrusted prompt inputs.
What do I get? / Deliverables
Incoming prompts and outgoing model text pass through defensive checks and hardening hooks so you can ship agent features with a clearer security baseline.
- Injection detection and output scanning hooks in the agent pipeline
- Prompt hardening patterns applied before high-risk runs
- Clearer pre-launch security posture for MCP-based agents
Recommended MCP Servers
Journey fit
How it compares
Agent-security MCP middleware—not a vulnerability scanner for traditional web apps or a generic code review skill.
Common Questions / FAQ
Who is agent-immune for?
Solo builders and small teams running AI agents via MCP who worry about prompt injection, unsafe outputs, and weak system prompts.
When should I use agent-immune?
Use it in Ship before exposing agents to customers or broad internal users, especially when inputs include web, email, or third-party documents.
How do I add agent-immune to my agent?
Install the agent-immune package from PyPI, register it as a stdio MCP server in Claude Code or Cursor, and wire your agent workflow to call its security tools around untrusted I/O.