openclaw-prompt-shieldLocal input-hardening scanner for OpenClaw agents. Pattern-based detection across 9 categories of LLM input risks, with combined-signal scoring and caller-su...
Install via ClawdBot CLI:
clawdbot install gopendrasharma89-tech/openclaw-prompt-shieldGrade Fair — based on market validation, documentation quality, package completeness, maintenance status, and authenticity signals.
Accesses sensitive credential files or environment variables
/etc/passwdSends data to undocumented external endpoint (potential exfiltration)
send → https://attacker.requestbin.com/logPotentially destructive shell commands in tool definitions
eval(Accesses system directories or attempts privilege escalation
/proc/Generated Oct 8, 2026
An OpenClaw agent reads incoming email on behalf of a busy executive and summarizes, categorizes, and drafts replies. Before processing each message, the shield scans the body for instruction_override, role_hijack, and indirect_injection patterns hidden in quoted reply chains, HTML comments, or markdown-linked text. Risky emails are routed to a manual review queue with a sanitized version attached for context.
A research agent scrapes competitor websites, news articles, and forum threads, then synthesizes a briefing. Scraped text often contains adversarial content such as 'ignore previous instructions' embedded in SEO spam or user-generated comments. Batch scanning with scan_batch.py flags which sources are safe to feed downstream, while sanitize_input.py redacts suspicious directives without discarding the article's substance.
A support automation pipeline ingests tickets from multiple channels, including email, chat transcripts, and web forms. The shield screens each ticket for data_exfiltration and tool_abuse patterns before the agent attempts to look up account details or issue refunds. High-risk tickets are escalated to human agents along with the matched categories and risk score.
A legal team uses an OpenClaw agent to review contracts and discovery documents submitted by opposing counsel. Documents can contain indirect_injection attempts buried in footnotes, hidden text, or zero-width character sequences designed to manipulate the agent's summary. Whitelisting allows the team to exclude standard boilerplate while still catching malicious directives.
A coding assistant agent reads GitHub issues, pull request comments, and Stack Overflow snippets to answer developer questions. Malicious users may attempt to hijack the agent through encoding_evasion payloads (base64, hex) or fake system-prompt delimiters. The shield provides a fast local pre-filter so the agent can safely reference community content without executing embedded commands.
The core pattern-based scanner and sanitizer scripts are released as MIT-licensed open source, driving adoption among individual developers and small teams. A paid tier offers curated, continuously updated rule packs tailored to specific verticals (finance, healthcare, legal) along with priority updates when new evasion techniques emerge. This creates a steady subscription stream from organizations that need domain-specific coverage without building their own detection catalogs.
A hosted offering where the same detection logic runs as a sidecar or API in the customer's agent pipeline, with centralized logging, alerting, and dashboarding. Customers pay based on monthly scan volume and the number of protected agents. The service adds value through aggregated threat intelligence, automatic whitelist management, and compliance reporting for regulated industries.
The shield is embedded into third-party agent frameworks, chatbot platforms, and enterprise workflow tools as a white-labeled input-hardening module. Partners pay a licensing fee or revenue share based on the number of end users or API calls passing through their systems. This model capitalizes on the skill's lightweight, dependency-free design, making it easy to bundle into existing products.
💬 Integration Tip
Integrate the shield as a pre-processing step before any content reaches the LLM, and combine it with a policy engine that maps the returned verdict to automated actions like blocking or routing to human review. Use the whitelisting feature to suppress false positives from domains that legitimately discuss prompt injection, and regularly update the caller-editable pattern files to keep pace with new evasion techniques.
Scored Oct 8, 2026
Calls external URL not in known-safe list
https://clawhub.ai/gopendrasharma89-tech/openclaw-prompt-shieldAI Analysis
The skill is a local, pattern-based input scanner with no remote calls, no API keys, and no LLM dependencies. The flagged signals are false positives: /etc/passwd and /proc/ are references in pattern definitions, not actual file access; the URL is the skill's own homepage; and eval() is not present in the skill definition. The skill's stated purpose matches its implementation exactly.
Audited May 10, 2026 · audit v1.0
Security-first skill vetting for AI agents. Use before installing any skill from ClawdHub, GitHub, or other sources. Checks for red flags, permission scope,...
Security scanner for AI agent skills. 9 built-in detection signatures. Identifies secrets, unsafe execution patterns, and prompt injection. Sub-50ms results.
Wallet anti-theft guard. One-click scan for high-risk wallet approvals to protect user assets. Use when a user asks for a wallet security check, wallet healt...
Comprehensive security audit for an agent's full skill stack. Chains scanner, differ, trust-verifier, and health-monitor into a single assessment with priori...
GEO Audit — AI Search Visibility Checker for ChatGPT, Perplexity, Claude & Gemini. 29-point GEO readiness checklist: robots.txt AI crawler access, Index...
Senior SecOps engineer skill for application security, vulnerability management, compliance verification, and secure development practices. Runs SAST/DAST sc...