AI Security Atlas › All entries › LLM Application & Output Security
Prompt Hygiene & Canary Detection
System prompts carry no secrets, include canary tokens, and outputs are scanned for prompt echoes.
Open in the interactive atlas →Impact
Turns prompt leakage into a detectable, low-impact event.
Why this severity
Guardrails take the highest severity of the risks they mitigate: Secrets & Policy in System Prompt (Medium), System Prompt Extraction (Medium).
Method: OWASP Risk Rating Methodology
Related
Mitigates
- Secrets & Policy in System Prompt Vulnerability · Medium severity
- System Prompt Extraction Threat · Medium severity
Builds on
- Input/Output Safety Classifiers Guardrail · Fixes high risks
Remediation
- Insert a unique canary string per deployment.
- Block or flag completions with high overlap to the system prompt.
- Review prompt templates in code review like source code.
How to verify
- Canary appearance raises a SIEM alert within one minute.
- Overlap detector blocks verbatim prompt dumps.
Pattern
if (completion.includes(CANARY) || overlap(completion, SYSTEM) > 0.6) return block();Framework mapping
OWASP Top 10 for LLM Applications 2025
- LLM07:2025 System Prompt Leakage
MITRE ATLAS
- AML.M0020 Generative AI Guardrails
NIST AI RMF 1.0
- MEASURE 2.4 AI system functionality and behavior are monitored in production