A lightweight two-stage safety framework for AI agents to filter prompts and prevent injection attacks before response generation.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install input-safety-guard
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install input-safety-guard using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Input Safety Guard?
Input Safety Guard is a specialized security framework within the Openclaw Skills ecosystem designed to act as a robust gatekeeper for AI agents. It provides a sophisticated two-stage screening process that protects against prompt injection, prompt leakage, and various categories of unsafe or malicious user input. By integrating this skill, developers can ensure their agents remain within safe operational boundaries without sacrificing performance or user experience.
The system leverages a unique combination of deterministic prefiltering and agent-native semantic review. This layered defense strategy allows for high-speed rejection of known threats while utilizing the LLM's own reasoning capabilities to understand complex intent. As a core component of Openclaw Skills, it helps maintain the integrity of agent instructions and prevents unauthorized data exposure in production environments.
Input Safety Guard Use Cases
- Blocking malicious prompt injection attempts in public-facing AI agents.
- Preventing sensitive system instructions and internal logic from being leaked to users.
- Automatically classifying and filtering risky requests such as hate speech, illegal activities, or PII extraction.
- Implementing customized safety strictness levels for different deployment environments using Openclaw Skills.
How Input Safety Guard Works
- The system intercepts the raw user message before any downstream prompt construction or processing occurs.
- Stage 1 Prefilter runs deterministic rules to normalize input and check against allowlists or known injection patterns.
- If Stage 1 returns a block status, the workflow stops immediately and returns a safe interception response.
- If Stage 1 permits, Stage 2 performs a semantic review using the agent's native reasoning to classify the request into specific unsafe families.
- The system validates the final decision based on the active profile; only requests with an allow decision are passed to the agent for a final response.
Input Safety Guard Setup
To integrate this into your workflow, ensure the following files are in your project directory: config/default_rules.yaml and the src/input_safety_guard/ source files. Use the following steps to initialize:
# Install required dependencies
pip install pyyaml
# Configure your rules in the yaml files
# default_rules.yaml, default_rules.strict.yaml, or default_rules.relaxed.yaml
Within your Python code, initialize the pipeline:
from input_safety_guard.pipeline import InputSafetyPipeline
pipeline = InputSafetyPipeline(profile="default")
response = pipeline.handle_user_message(user_input)
Input Safety Guard Data Schema & Taxonomy
The Input Safety Guard organizes its decisions into structured metadata for clear auditing and routing. The primary schema includes:
| Field | Type | Description |
|---|---|---|
| decision | String | The final safety result: allow, review, or block. |
| source | String | Indicates if the decision came from prefilter or stage2. |
| category | String | The specific risk classification (e.g., prompt_leaking, insult). |
| confidence | String | Reliability of the classification: high, medium, or low. |
| matched_rules | Array | List of deterministic rules triggered during Stage 1. |
Input Safety Guard Advanced Features
- Multi-profile support (strict, relaxed, default) allowing developers to tune Openclaw Skills behavior based on trust levels.
- Canonical Stage 2 prompt templates to ensure consistent semantic analysis across different agent instances.
- Support for structured metadata output, enabling deep integration with logging and observability platforms.
- Deterministic pre-filtering for high-speed mitigation of known threats without calling the LLM API unnecessarily.
SKILL.md
Loading
Related Openclaw Skills
HF Daily Papers Digest
A streamlined automation skill to fetch, analyze, and generate professional digests from the HuggingFace Daily Papers feed.
git-quick
A lightweight Git intelligence tool that provides immediate visibility into repository status, commit history, and team contribution metrics.
Cyber Palace AI Roleplay
A fully autonomous AI agent skill that navigates a cyber-palace environment, making strategic social and political moves every two minutes.
Claude Tamagotchi
A real-time virtual pet system for AI agents and developers featuring evolving pixel art species and high-stakes survival mechanics.
Session Management
An integrated suite for managing the full lifecycle of Claude Code sessions, from keyword searching to automated history repair.
TikTok Ad Library & Creative Discovery (PipiAds)
An AI agent skill for discovering and analyzing high-performing TikTok ad creatives and marketing patterns via the PipiAds database.
Featured*
