Suspicious AI Dataset Label Modification

Synopsis

Field	Value
Activation Period	14 Days
Training Period	30 Days
Test Period	N/A (single event)
Deduplication Period	5 Days
Required Data	Requires one of the following data sources: AWS Audit Log OR Azure Audit Log OR Gcp Audit Log
Detection Modules	Cloud
Detector Tags	Cloud AI Infrastructure Analytics
ATT&CK Tactic	Impact (TA0040)
ATT&CK Technique	Data Manipulation: Stored Data Manipulation (T1565.001)
Severity	Low

Description

AI Dataset labels were modified by an identity that typically doesn't interact with labels. MITRE ATLAS Technique: AML.T0020 - Poison Training Data. OWASP Top 10 LLM Technique: LLM04 - Data and Model Poisoning.

Attacker's Goals

Contaminating training set, so that predictions on new data will be modified.

Investigative actions

• Check the identity that modified the dataset's labels.
• Check that the dataset is correctly labeled.