Guard System plugin thumbnail

Guard System

Security guards for tool execution and prompt safety. Blocks tools with flagged scan status and detects prompt injection patterns.

Author PaoloC68 1 stars Version 1.0.0 Updated
SecurityToolsPrompting
GitHub Repository Discussion How to install

README

Guard System

Security guards for Agent Zero tool execution and prompt safety. Blocks tools associated with flagged scan results and detects prompt injection patterns in assembled prompts.

Features

  • Scan Status Guard — Blocks tool execution when a skill/tool has been flagged as blocked by an external security scanner. Tools flagged as needs_review produce a warning but are allowed to run.
  • Prompt Length Guard — Monitors assembled prompt size and warns the LLM when content exceeds a configurable character limit.
  • Injection Detection — Scans prompts for known injection patterns (regex-supported) and surfaces warnings to the LLM context.

Configuration

Option Type Default Description
enabled bool true Master switch for all guards
max_prompt_length int 500000 Maximum allowed prompt length in characters
injection_detection bool true Enable prompt injection pattern scanning
injection_patterns list[str] (see default_config.yaml) Regex patterns to detect injection attempts
scan_results dict {} Tool/skill scan results from external scanners

How Scan Results Work

External scanners (or manual review) write results into the plugin's scan_results config. Each key is a tool or skill name, and the value is a dict with status and optional metadata.

Statuses

Status Behavior
blocked Tool execution is prevented. A RepairableException is raised so the LLM can try a different approach.
needs_review A warning is printed but execution proceeds.
(anything else) No action taken.

Setting Scan Results via Python

from helpers.plugins import get_plugin_config, save_plugin_config

config = get_plugin_config("guard-system", agent=agent) or {}
config.setdefault("scan_results", {})["my_skill"] = {
    "status": "blocked",
    "reason": "Malicious code detected",
    "scanner": "cisco"
}
save_plugin_config("guard-system", settings=config)

Clearing a Scan Result

config = get_plugin_config("guard-system", agent=agent) or {}
config.get("scan_results", {}).pop("my_skill", None)
save_plugin_config("guard-system", settings=config)

Extension Points

Hook File Purpose
tool_execute_before _05_scan_status_guard.py Checks scan results before each tool runs
message_loop_prompts_after _05_prompt_length_guard.py Checks prompt size and injection patterns after prompt assembly

Installation

Copy the guard-system/ directory into usr/plugins/ and enable it from the Agent Zero settings UI.