Prompt Injection Mitigation Guide

Attack Type Classification

Type	Objective	Example	Detection Indicator
Instruction Override	Role Destruction	"Ignore previous"	Role Deviation Rate
Data Extraction	Secret Information Elicitation	"Show system prompt"	Forbidden Token Occurrence
Context Poisoning	Control Flow Alteration	Malicious Embedding Injection	Anomalous Similarity Score
Chain Reversal	Guard Evasion	Meta-instructions / Self-reference	Deviation Chain Length

Defense-in-Depth Model

1. Input Sanitization (control tokens / URLs / Base64)
2. Malicious candidate filtering via embedding similarity
3. Two-stage generation (intent summary → approval → actual generation)
4. Output policy filter (regex + classification model)
5. Audit logging (prompthash, decisionreason)

Validation Metrics

Metric	Definition	Target
False Positive Rate	Benign blocked	< 3%
False Negative Rate	Malicious passed	< 5%
Average Latency Increase	Defense-added delay	< 300ms

Next Actions

• Automate test corpus generation script
• Document malicious signature update procedures

Back to: index.md