prompt-guard

name	prompt-guard
description	Detects prompt injection patterns in graph content with risk classification
type	semantic
user-invocable	true
semantic	{"skillVersion":1,"permissions":{"graph":["read","write"],"proofs":["request"],"memory":["recall"]},"assertions":[{"predicate":"guard:injection-detected","requireProof":true},{"predicate":"guard:scan-result","requireProof":false}],"queries":[{"predicate":"guard:history","scope":"agent"}]}

prompt-guard

Detects prompt injection patterns in graph content and classifies findings by risk level.

When to Use

Use this skill when:

• Inspecting user-submitted text stored in the knowledge graph for injection attempts
• Auditing graph content before it reaches downstream agents or tools
• Monitoring for adversarial inputs such as role overrides, encoding evasion, or jailbreak prompts

Risk Classification

Each finding is classified into one of three levels:

• dangerous -- Active injection attempts (role overrides, system overrides, jailbreak, shell commands)
• suspicious -- Evasion techniques or anomalies (zero-width characters, homoglyphs, encoding tricks, template injection)
• safe -- No injection patterns detected

The overall classification for a scanned item is dangerous if any dangerous finding exists, suspicious if only suspicious findings exist, and safe otherwise.

Instructions

1. Store text content in the graph using skill_assert or skill_graph_query
2. Query with predicate guard:scan-result to trigger the prompt-guard runtime
3. The skill enriches each result with a classification field and a findings array
4. Use guard:injection-detected (with proof) to record confirmed injection attempts
5. Query guard:history to review past scan results for the current agent