Adaptive security toolkit for AI agents. Assess inputs for prompt injection, scan outputs for credential/PII leaks, teach new attack patterns to semantic memory, harden prompts with role-lock and sandboxing, and monitor metrics — all locally via MCP. Zero false positives on benchmarks, 179 tests, Apache-2.0.
Overview
agent-immune
Adaptive security toolkit for AI agents — runs locally via MCP.
Tools
- assess_input — score text for prompt injection threats
- assess_output — scan output for credentials, PII, and exfiltration
- learn_threat — teach new attack patterns to semantic memory
- harden_prompt — apply role-lock, sandboxing, and output guards
- get_metrics — observability snapshot (counts, latency)
Install
pip install 'agent-immune[mcp]' python -m agent_immune serve --transport stdio
Highlights
- Zero false positives on benchmarks
- 179 tests, 94% coverage
- Semantic memory catches rephrased attacks that bypass regex
- Apache-2.0 license
Server Config
{
"mcpServers": {
"agent-immune": {
"command": "python",
"args": [
"-m",
"agent_immune",
"serve",
"--transport",
"stdio"
]
}
}
}