About this skill

Problem

When LLMs are embedded in production workflows, security review is often implicit: user messages may expose ID cards, bank cards, internal IPs, cloud metadata endpoints, or exfiltration webhooks; skill installation and execution may still be bypassed by prompt injection, path traversal, encoded payloads, or credential access. Relying on model self-checks is hard to audit and weak for enterprise compliance.

How It Works

AIDR XClaw Security Sentinel makes the control path explicit through two gates:
- Gate 1 Query Audit: each user message is locally desensitized first, replacing ID card, phone, API key, Bearer token, and SSRF target patterns with placeholders, then an audit API decides pass, warn, or block.
- Gate 2 Skill Audit: before installing or running a skill, it generates L2/L3/final fingerprints, packages from a temporary directory, uploads for Pre-Install or Runtime audit, and requires staged outputs plus a full report.

Boundaries

API failures must be treated as blocking and must not be fabricated. Query Audit must not send raw sensitive content to the cloud. It fits AI-ops environments that need a traceable audit trail, not offline deployments, environments without audit services, or cases intended to bypass security controls.

Screenshots

Use Cases

Before serving user messages, locally redact sensitive fields and call the audit API to pass, warn, or block requests.
Before installing third-party skills, package from a temp directory, generate fingerprints, and run Pre-Install audit.
When running installed skills, redact content before upload, execute Runtime audit, and emit staged reports.
During prompt-risk review, follow the API safety_level to choose pass, warn, or block and retain the full report.

Best For

AI gateway security engineers who need to redact user messages and decide pass, warn, or block before business logic.
Agent-platform ops leads who need to verify skill content, fingerprints, and install-time risk before release.
Architecture owners who need to include prompt-injection detection in a traceable large-model compliance workflow.
Internal tooling teams who need to stop runtime skills from reading sensitive configs or calling exfiltration webhooks.