Skip to content
USE CASE: OPENCLAW

Secure OpenClaw agents

OpenClaw agents execute shell commands, browse the web, and message across 15+ channels. PromptGuard scans every prompt, tool call, and response to prevent exploitation of CVE-2026-25253, CVE-2026-32918, and indirect prompt injection.

Start for FreeView documentation
THREE EXITS, THREE EXPOSURESschematic
the agentshell execdestructive command patternsblockedweb fetchinjected directives in fetched contentblockedoutbound messagepii · exfiltrationredacted
An agent that can run a shell, read the web and message a channel has three different exposures, not one. Each exit is checked against the thing that can actually go wrong there.

Where an OpenClaw agent is exposed

  1. 01

    Indirect Prompt Injection Defense

    OpenClaw agents process web content, files, and messages from untrusted sources. PromptGuard's six-layer detection catches hidden instructions before they reach the model.

  2. 02

    Tool Call Validation

    Every shell command, file operation, and API call is validated against security policies before execution. Block privilege escalation and data exfiltration attempts.

  3. 03

    Multi-Turn Drift Detection

    Crescendo attacks across conversation turns are caught by semantic embedding analysis. Even if each message looks safe, PromptGuard detects when the trajectory turns harmful.

  4. 04

    Multi-Channel PII Protection

    OpenClaw sends messages to WhatsApp, Telegram, Slack, and Discord. PromptGuard redacts PII before it reaches any channel, preventing data leaks across messaging platforms.

  5. 05

    Session Isolation Enforcement

    Detect attempts to escape sandbox boundaries (CVE-2026-32918). PromptGuard monitors for cross-session data access patterns.

  6. 06

    Skill Marketplace Scanning

    341 malicious skills were found in OpenClaw's public marketplace. PromptGuard scans skill instructions for embedded prompt injections and malicious tool schemas.

How it works for OpenClaw

  1. 01

    Route through PromptGuard

    Point OpenClaw's LLM configuration to PromptGuard's proxy. One config change, zero code changes.

  2. 02

    Every call is scanned

    Six detection layers analyze every prompt and tool call: normalization, regex, ML ensemble, content safety, multi-turn drift, and policy evaluation.

  3. 03

    Threats are blocked

    Malicious prompts are blocked before reaching the LLM. PII is redacted before leaving through messaging channels. Every action is logged for audit.

Secure OpenClaw in one line

python
import promptguard
from openclaw import OpenClaw

# Auto-instruments all LLM calls made by OpenClaw
promptguard.init(
    api_key="your-api-key",
    # PromptGuard proxies and secures every request
    # No changes needed to OpenClaw's agent code
)

agent = OpenClaw(
    model="gpt-5-nano",
    skills=["web-browse", "shell", "messaging"],
    channels=["slack", "telegram", "discord"],
)

# All prompts, tool calls, and responses are now
# scanned by PromptGuard's six detection layers
agent.run("Summarize the latest security advisories")

The controls this sector runs on

Secure your OpenClaw agents today

Protect OpenClaw agents from prompt injection, data exfiltration, and sandbox escapes. Enterprise-grade security with zero code changes.

Get started freeContact sales