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

GuardrailFixes high risks

Dependency Allowlisting & Package Verification

New packages must come from an approved registry or mirror and pass provenance, age, and malware checks before install.

Impact

Hallucinated or typo-squatted names cannot be installed by people or agents.

Why this severity

Guardrails take the highest severity of the risks they mitigate: Unreviewed AI-Suggested Code & Dependencies (High), Hallucinated Package Squatting (High).

Method: OWASP Risk Rating Methodology

Related

Mitigates

• Unreviewed AI-Suggested Code & Dependencies Vulnerability · High severity
• Hallucinated Package Squatting Threat · High severity

Remediation

1. Mirror approved packages internally and point package managers at the mirror.
2. Require lockfiles with hashes.
3. Scan new packages for install scripts and known malware.

How to verify

• Installing an unknown package from a fresh checkout fails.
• Lockfile hash mismatches fail the build.

Pattern

npm config set registry https://registry.internal.example/
npm ci --ignore-scripts

Framework mapping

OWASP Top 10 for LLM Applications 2025

• LLM03:2025 Supply Chain
• LLM09:2025 Misinformation

MITRE ATLAS

• AML.M0016 Vulnerability Scanning
• AML.M0011 Restrict Library Loading

NIST AI RMF 1.0

• MANAGE 3.1 Third-party AI risks are monitored and controls applied