seojoonkim/prompt-guard

Overview

This skill is an advanced prompt injection defense engine that runs fully offline with 577+ bundled patterns and optional API access for early-access and premium rules. It performs multi-language detection across 10 languages, assigns severity scores, and outputs actionable results such as block, warn, or log. Tiered pattern loading and a hash-backed cache make scans efficient for real-time agent use.

How this skill works

The engine normalizes input, applies a tiered pattern matcher, and scores findings across 11 SHIELD categories (e.g., prompt, tool, memory, supply_chain). Results include severity, recommended action, matched patterns, and a SHA-256 fingerprint for deduplication. Optional API can augment local patterns with early-access or premium detections while a local cache reduces repeated work.

When to use it

• Protect an LLM-based assistant from jailbreaks and instruction overrides
• Scan user inputs and LLM outputs before execution or tool invocation
• Harden agent tooling that exposes system commands, files, or network access
• Audit skill/plugin requests for exfiltration, key injection, or obfuscated payloads
• Integrate into CI or security pipelines to detect prompt supply-chain risks

Best practices

• Start with the default HIGH tier and CRITICAL always enabled; escalate to MEDIUM or FULL on suspicious inputs
• Enable the local cache to avoid repeated pattern re-evaluation and improve latency
• Use sanitize_output() on model responses destined for downstream tools or external systems
• Configure actions per severity to match operational workflow (log, warn, block, notify)
• Keep API disabled for fully offline deployments; enable only when you need early-access or premium patterns

Example use cases

• Pre-flight check of LLM tool calls that execute shell commands or create files
• Enterprise agent gateway that blocks secrets exfiltration and SSH key injection attempts
• Security auditing of third-party skills/plugins for policy_bypass and supply_chain threats
• Automated monitoring for semantic worms, reverse shells, and obfuscated payloads in chat logs
• Tiered scanning in resource-constrained environments: CRITICAL quick-scan then escalate if needed