AI Security Atlas › All entries › Model Supply Chain & Training Integrity

GuardrailFixes high risks

Dataset Provenance & Poison Screening

Training data is traced to source, filtered for anomalies, and screened for trigger patterns before each run.

Impact

Reduces the chance that a backdoor survives into a released checkpoint.

Why this severity

Guardrails take the highest severity of the risks they mitigate: Training & Fine-Tune Data Poisoning (High), Unverified Data & Model Provenance (High).

Method: OWASP Risk Rating Methodology

Related

Mitigates

• Training & Fine-Tune Data Poisoning Threat · High severity
• Unverified Data & Model Provenance Vulnerability · High severity

Needed by

• PII Scrubbing & Privacy-Preserving Training Guardrail · Fixes high risks

Remediation

1. Hash and log every shard with its origin and license.
2. Apply spectral-signature or activation-clustering screens on fine-tune sets.
3. Gate releases on behavioral-diff evals against the prior model.

How to verify

• Planted poison fixtures are flagged before training starts.
• Release checklist links the provenance report for each dataset.

Framework mapping

OWASP Top 10 for LLM Applications 2025

• LLM04:2025 Data and Model Poisoning

MITRE ATLAS

• AML.M0007 Sanitize Training Data
• AML.M0025 Maintain AI Dataset Provenance

NIST AI RMF 1.0

• MAP 4.1 Risks of AI components, including third-party data and software, are mapped