AI Security Atlas › All entries › Model Supply Chain & Training Integrity
Inference Quotas & Output Watermarking
Per-tenant quotas, scoped access to sensitive outputs, and watermarking deter extraction and support attribution.
Open in the interactive atlas →Impact
Raises the cost of model theft and makes distilled copies traceable.
Why this severity
Guardrails take the highest severity of the risks they mitigate: Unrestricted Inference Endpoints (High), Model Extraction via Inference API (High).
Method: OWASP Risk Rating Methodology
Related
Mitigates
- Unrestricted Inference Endpoints Vulnerability · High severity
- Model Extraction via Inference API Threat · High severity
Remediation
- Token-bucket limits per key and per organization.
- Disable logprobs unless the caller holds a research scope.
- Embed statistical watermarks in generated text where supported.
How to verify
- Extraction simulations hit quota before reaching useful fidelity.
- Watermark detector identifies outputs from a sampled corpus.
Pattern
if (!bucket.take(key, tokens)) return res.status(429).end();Framework mapping
OWASP Top 10 for LLM Applications 2025
- LLM10:2025 Unbounded Consumption
MITRE ATLAS
- AML.M0004 Limit AI Service Query Volume and Rate
NIST AI RMF 1.0
- MEASURE 2.7 AI system security and resilience are evaluated and documented