AI Security Atlas › All entries › LLM Application & Output Security
Input/Output Safety Classifiers
Independent classifiers screen prompts and completions against application and vendor policy.
Open in the interactive atlas →Impact
Adds a second, differently trained layer that jailbreaks must beat.
Why this severity
Guardrails take the highest severity of the risks they mitigate: Single-Layer Safety Reliance (Medium), Jailbreak & Safety Bypass (High).
Method: OWASP Risk Rating Methodology
Related
Mitigates
- Single-Layer Safety Reliance Vulnerability · Medium severity
- Jailbreak & Safety Bypass Threat · High severity
Needed by
- Prompt Hygiene & Canary Detection Guardrail · Fixes medium risks
Remediation
- Run input screening before the main model call.
- Run output screening before streaming to the user.
- Tune thresholds per surface and audience.
How to verify
- False-negative rate measured on a labeled jailbreak set each release.
- Classifier latency stays within the streaming budget.
Framework mapping
OWASP Top 10 for LLM Applications 2025
- LLM01:2025 Prompt Injection
MITRE ATLAS
- AML.M0020 Generative AI Guardrails
NIST AI RMF 1.0
- MEASURE 2.6 AI system is evaluated regularly for safety risks