Training dataset without integrity record

ID

training-dataset-without-integrity-record

Severity

high (catalogue Medium; platform severity model has no 'medium')

Remediation Complexity

trivial

Remediation Risk

low

Remediation Effort

low

OWASP LLM

LLM04:2025 — Data and Model Poisoning

OWASP ASI

ASI06:2026 (secondary ASI04:2026)

Family

LLM04 — Data and Model Poisoning

Red-team vectors

RAG Poisoning

Tags

ai_security

Description

A dataset used for training / fine-tuning has no integrity record — no checksum, content hash, version, or revision pin. Without one, a poisoned dataset swapped in upstream is undetectable: the model trains on tampered data with no signal. The training-purpose leg is required; an evaluation dataset without integrity is out of scope, and any integrity signal in the file clears the finding.

Examples

train_dataset = load_dataset("s3://bucket/corpus")   (1)
trainer = Trainer(model=model, train_dataset=train_dataset)
1 Training data pulled with no checksum / version pin — flagged.

Recording a sha256 / content hash or pinning a dataset revision produces no finding.

Mitigation / Fix

  • Record and verify a dataset integrity value (checksum / content hash) before training.

  • Pin a dataset version / revision so replacements are detectable.