A Chromium extension that sanitizes web pages by masking PII, blocking dark patterns, and neutralizing prompt injections before an AI agent parses them.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install agent-browser-shield
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 agent-browser-shield 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 Agent Browser Shield?
Agent Browser Shield is a specialized Chromium MV3 extension designed to act as a proactive sanitization and security layer for autonomous web agents. Running directly inside the browser instance, it processes web pages before your agent parses the DOM. By masking personally identifiable information (PII), neutralizing manipulative dark patterns, blocking hidden cart upsells, and stripping dangerous prompt-injection surfaces, the tool ensures high-integrity automated operations. Developers looking to implement robust safety protocols across their Openclaw Skills can use this extension to minimize operational risks and restrict LLM vulnerability to untrusted web content.
Agent Browser Shield Use Cases
- Safeguarding automated workflows from indirect prompt injection vulnerabilities embedded in unverified web pages.
- Preventing AI web agents from leaking sensitive user credentials or critical database keys during standard processing.
- Neutralizing deceptive checkout patterns, auto-checked upcharges, and sneaky cart additions during automated purchasing runs.
- Reducing token costs and context window noise by permanently hiding ads, tracking cookies, and scarcity pop-ups.
How Agent Browser Shield Works
- Initialize the browser profile containing the extension via local headless/headed execution or remote browser instances.
- The extension automatically hooks into the tab loading cycle, processing the DOM before it is exposed to the runner.
- Dynamic content filters sanitize sensitive elements, stripping tracking code and normalizing coercive confirm-shaming dialogues.
- The sanitized elements are flagged with specialized CSS classes and data attributes to maintain structural context for the agent.
- The AI agent processes the clean DOM while adhering to required behavior protocols outlined by the injected markers.
Agent Browser Shield Setup
Path 1: Local Chrome via Remote CDP
Launch your local Chromium browser with remote debugging enabled and load the unpacked extension archive:
chrome --remote-debugging-port=9222 --user-data-dir=/path/to/profile --load-extension=/path/to/unpacked-extension --disable-extensions-except=/path/to/unpacked-extension
Register the newly created profile with your local runner to configure your Openclaw Skills environment:
openclaw browser create-profile --name shield --driver remote --cdp-url http://127.0.0.1:9222
openclaw browser use shield
Path 2: Browserbase Integration
Upload the extension ZIP file to Browserbase to secure cloud-hosted execution:
openclaw plugins install @browserbasehq/openclaw-browserbase
openclaw browserbase setup
npm install -g browse
export BROWSERBASE_API_KEY="your_api_key_here"
browse cloud extensions upload ./agent-browser-shield-extension.zip
Save the printed extension ID into memory and reference it during Browserbase runner instantiation to maintain performance.
Agent Browser Shield Data Schema & Taxonomy
Element Metadata and DOM Taxonomy
To ensure proper state synchronization, Agent Browser Shield marks sanitized elements using custom data attributes. Agents should consult this reference schema:
| Selector | Description | Expected Behavioral Guardrail |
|---|---|---|
[data-abs-rule="<rule-id>"] |
Any element mutated by an active shielding rule. | Inspect structural modifications during trace debugging. |
.abs-placeholder |
Interactive placeholder replacing hidden secrets or raw PII. | Do not expand or parse unless explicitly requested by the user. |
.abs-cart-addon-annotate |
Visual flag highlighting pre-added shopping cart upsells. | Report the pre-added item to the user; do not interact or validate blindly. |
data-abs-cleared |
Denotes a pre-checked input field that has been force-unchecked. | Leave unchecked unless explicitly commanded to opt-in. |
data-abs-confirmshame-orig-text |
Preserves rewritten texts that originally used guilt-tripping phrasing. | Safe to trigger normally. Original phrasing is preserved in attribute. |
Agent Browser Shield Advanced Features
- Granular rules toggling: Access the embedded options control panel to selectively enable or disable specific privacy protections.
- Scalable cloud agent support: Easily upload the pre-built bundle into serverless browser clusters to support remote Openclaw Skills.
- Safe-state preservation: Preserves critical functional DOM elements while rewriting aggressive visual or textual prompts to ensure seamless scripting.
- Adaptive target filtering: Dynamically intercepts malicious prompt structures natively, preventing malicious page inputs from overriding original system instructions.
SKILL.md
Loading
Related Openclaw Skills
Expense Tracker Agent Skill
A scheduled, AI-driven daily expense tracker that processes Telegram check-ins, deduplicates entries, and syncs local Markdown ledgers with Google Sheets.
8D Pattern Awareness
This skill transforms unstructured health logs, daily check-ins, or voice transcripts into structured, non-diagnostic wellness pattern maps using the 8 dimensions of wellness.
Cloudflare Developer Platform Skill
A comprehensive reference guide and action plan for building, deploying, and optimizing serverless workflows on the Cloudflare Developer Platform.
Find Software Developer
An AI agent skill to search, filter, and enrich information on over 100k+ US software development firms via the ServiceGraph API.
Model Configuration and Routing
A technical capability designed to configure, switch, and troubleshoot model providers, routing sessions, and cron executions in OpenClaw.
MCPorter
MCPorter is a versatile command-line interface designed to configure, authenticate, inspect, and call Model Context Protocol (MCP) servers and tools directly via HTTP or stdio.
Featured*
