ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Watchers (aka The Norton for OpenClaw) # 🦞🛡️ ClawKeeper: Comprehensive Safety Protection for OpenClaw Agents Through Skills, Plugins, and Watchers

(aka The Norton for OpenClaw)

SAFETY EXFOLIATE! SAFETY EXFOLIATE!

**ClawKeeper** is a _comprehensive real-time security framework_ designed for autonomous agent systems such as **OpenClaw**. It provides unified protection through three complementary approaches: **skill-based** safeguards at the instruction level, **plugin-based** enforcement at the runtime level, and a **watcher-based** independent monitoring agent for external oversight. # 🔎 Overview **ClawKeeper** provides protection mechanisms across three complementary architectural layers: - **Skill-based Protection** operates at the instruction level, injecting structured security policies directly into the agent context to enforce environment-specific constraints and cross-platform boundaries. - **Plugin-based Protection** serves as an internal runtime enforcer, providing configuration hardening, proactive threat detection, and continuous behavioral monitoring throughout the execution pipeline. - **Watcher-based Protection** introduces a novel, decoupled system-level security middleware that continuously verifies agent state evolution. It enables real-time execution intervention without coupling to the agent's internal logic, supporting operations such as halting high-risk actions or enforcing human confirmation. Importantly, **Watcher-based Protection** is **system-agnostic** and can be integrated with different agent platforms to provide regulatory separation between task execution and safety enforcement, enabling **proactive** and **adaptive** security across the entire agent lifecycle. It can be deployed both **locally** and in the **cloud**, supporting personal deployments as well as enterprise or intranet environments. ![](fig/overview.png) ![](fig/item_list.png) # 📦 Installation ClawKeeper supports three complementary protection mechanisms. ## 📚 I. [Skill-based Protection](clawkeeper-skill/README.md) Inject security policies directly into the agent context through structured Markdown documents and scripts. **Quick Start:** ### Windows Safety Guide ```powershell cd clawkeeper-skill/skills/windows-safety-guide ./scripts/install.ps1 ``` Then instruct OpenClaw: ``` Please use the windows-safety-guide skill to enforce behavior security policies, configuration protection, and enable nightly security audits. ``` ### Feishu (Lark) Safety Guide ```bash cd clawkeeper-skill/skills/feishu-safety-guide bash scripts/install.sh ``` Then instruct OpenClaw: ``` Please use the feishu-safety-guide skill to enforce message protection, credential security, and enable periodic security reporting in Feishu (Lark). ``` For **detailed** setup options and deployment from prompt, see [Skill-based Protection](clawkeeper-skill/README.md). --- ## 📚 II. [Plugin-based Protection](clawkeeper-plugin/README.md) A runtime enforcer plugin providing configuration auditing, threat detection, and behavioral monitoring. **Quick Start:** ### Linux/macOS ```bash cd clawkeeper-plugin bash install.sh ``` ### Windows ```powershell cd clawkeeper-plugin ./install.ps1 ``` Then verify installation: ```bash npx openclaw clawkeeper audit ``` For **detailed** command reference and advanced usage, see [Plugin-based Protection](clawkeeper-plugin/README.md). --- ## 📚 III. [Watcher-based Protection](clawkeeper-watcher/README.md) An independent, decoupled governance layer providing runtime monitoring and execution control without coupling to the agent's internal logic. **Quick Start:** ### Prerequisites - Node.js and npm/pnpm installed - Git repository cloned ### Installation Steps 1. **Install repository dependencies:** ```bash pnpm install ``` 2. **Build and link the launcher:** ```bash cd clawkeeper npm install npm run build npm link cd .. ``` 3. **Initialize operating modes:** ```bash clawkeeper init remote clawkeeper init local ``` 4. **Launch the watcher:** ```bash # Remote governance mode clawkeeper remote gateway run # Local governance mode clawkeeper local gateway run ``` For **detailed** configuration, command reference, and feature documentation, see [Watcher-based Protection](clawkeeper-watcher/README.md). --- # 💡 Features ![](fig/framework.png) - **Comprehensive Security Scanning:** Regularly scans the runtime environment, dependencies, and workspace for vulnerabilities, providing clear and actionable risk alerts before threats occur. - **Real-time Threat Prevention & Gating:** Evaluates AI actions in real time, blocking high-risk behaviors such as prompt injection, credential leakage, and code injection. - **Behavioral Profiling & Anomaly Detection:** Builds long-term behavioral baselines for AI agents and detects anomalies when unusual actions, risky tool calls, or dangerous commands appear. - **Intent Enforcement & Trajectory Analysis:** Monitors multi-turn interactions to ensure AI actions stay aligned with the user’s original intent and prevents goal drift, unsafe loops, or unauthorized actions. - **Config Integrity & Drift Monitoring:** Protects critical configuration files and alerts users when unexpected changes weaken security settings or introduce new risks. - **Automated Hardening & Remediation:** Provides vulnerability remediation suggestions, applies secure default configurations, and supports one-click rollback with automatic backups. - **Third-Party Extension Shield:** Reviews and monitors external extensions and plugins to prevent malicious behavior or excessive permission access. - **Comprehensive Logging & Auditing:** Maintains full logs of user inputs, AI outputs, tool usage, and security decisions for auditing, compliance, and traceability. - **Self-Evolving Threat Intelligence:** Stores high-risk events and decisions to build a threat intelligence library that helps detect and prevent recurring or new attack patterns. - **Cross-Platform Ecosystem Security:** Ensures consistent security protection across operating systems and third-party platforms, providing full ecosystem coverage. --- # 🔬 Comparative Analysis of Safety Paradigms in ClawKeeper ClawKeeper offers a comprehensive suite of security mechanisms, allowing users to freely select and combine them according to their specific requirements, whether prioritizing runtime efficiency or security performance. ![](fig/compare.png) # 📈 Experiment Results To systematically assess the security capabilities of ClawKeeper, we construct a benchmark comprising seven categories of safety tasks, each containing 20 adversarial instances divided equally into 10 simple and 10 complex examples. We compare ClawKeeper against the most prominent open-source security repositories for OpenClaw-style agent ecosystems. The results showed that ClawKeeper achieved optimal defense performance. ![](fig/results.png) --- # 🔥 Updates - [2026-04-07] 🛡️ ClawKeeper v1.1 — new guard pipeline & security hardening: - **Execution Gate (`exec-gate`)**: Regex-based dangerous command detector that blocks destructive shell commands (e.g., `rm -rf /`, fork bombs, `curl | sh`, disk wipes) before agent execution. - **Path Guard (`path-guard`)**: Protected path enforcement that prevents agents from reading, writing, or deleting sensitive files (e.g., `~/.ssh/**`, `~/.aws/credentials`, `/etc/shadow`). - **Input Validator (`input-validator`)**: Lightweight JSON-Schema-subset validator that rejects malformed tool inputs (missing fields, wrong types, oversize strings, NUL bytes) at the interface boundary. - **Budget Guard (`budget-guard`)**: Rolling-window token budget control that halts agent execution when configured input/output/total token limits are exceeded. - **Permission Store (`permission-store`)**: Persistent allow/deny decisions keyed by (tool, fingerprint) with session and forever scopes, enabling operator-controlled authorization. - **CLI interface (`cli.js`)**: New `openclaw clawkeeper permission` commands for managing allow/deny rules from the command line. - **Tool schemas**: Added structured schemas for `bash`, `read_file`, and `write_file` tools. - **Security hardening**: Fail-closed policy enforcement, scoped permission bypass (allow no longer skips budget-guard and input-validator), and HMAC-SHA256 integrity protection for permission store files. - [2026-03-25] 🎉 ClawKeeper v1.0 has been released. - [2026-03-26] 🧠 We released our [paper](https://arxiv.org/abs/2603.24414) --- # 📝 License This project is licensed under [MIT](https://opensource.org/licenses/MIT). Source: https://github.com/SafeAI-Lab-X/ClawKeeper

@ClawKeeper

ClawKeeper