Your AI Assistant,
Protected
You use AI assistants like OpenClaw every day — reading files, sending messages, running tools. OpenGuardrails makes sure nothing goes wrong.
Real-World Example
How It Works with OpenClaw
OpenClaw is a personal AI assistant that acts on your behalf. Here's how OpenGuardrails protects every step.
Without Protection
- ✗Agent deletes wrong file
- ✗Sends email to wrong recipient
- ✗Runs malicious or unverified skill
- ✗Exposes sensitive data externally
- ✗Executes outside authorized scope
With OpenGuardrails
- ✓Actions checked before execution
- ✓Risky steps require human approval
- ✓Unsafe objects blocked automatically
- ✓Every execution logged and replayable
- ✓Workflows stay within boundaries
You ask OpenClaw to organize your project files
OpenClaw plans to move, rename, and delete files across multiple directories.
OpenGuardrails scans every object
Each file, directory, and action is checked against safety policies before execution.
Risky actions trigger approval
Deleting a folder? Sending an email? You get a confirmation prompt before it happens.
Everything is logged
Full audit trail of what happened, what was blocked, and what you approved.
Protection
Personal AI Assistant Protection
Six security capabilities that protect every action your personal AI agent takes.
scan_object()Object Trust
Judge whether a skill, plugin, file, link, webpage, email, or document is safe before your agent touches it.
check_action()Action Safety
Evaluate whether a delete, send, upload, or external write should be allowed before it happens.
require_approval()Human Approval
High-risk actions trigger confirmation. Your agent pauses, you decide, control stays with you.
bound_run()Bounded Execution
Limit task boundaries, tool chains, budgets, and time. Keep longer runs within safe bounds.
log_evidence()Evidence Logging
Record every input, risk judgment, policy hit, and final action. Full audit trail for every execution.
replay_run()Execution Replay
Replay any past execution step by step. See what happened, what was stopped, and debug with confidence.
Security Evaluation
Test Your AI Assistant's Safety
We simulate real-world attack patterns against your personal AI assistant to find vulnerabilities before attackers do.
Malicious Skill Injection
Test if your assistant can be tricked into running compromised or poisoned skills and plugins.
Indirect Prompt Injection
Attack through files, emails, webpages, and documents that your assistant processes.
Data Exfiltration
Test whether attackers can trick your assistant into leaking sensitive data to external services.
Dangerous Actions
Simulate destructive file, email, and shell operations your assistant might execute unchecked.
Context & Memory Poisoning
Test attacks that corrupt your assistant's context window or long-term memory.
Tool Misuse & Abuse
Evaluate whether MCP tools and integrations can be exploited beyond intended use.
Start Protecting Your AI Assistant
Free to start. Upgrade when your assistant becomes part of your daily life.