Design & Development

Training on Personal Data Without Consent

Backdoor Trigger Sample Implantation

Poisoning of the Retrieval Knowledge Base

Safety Alignment Broken During Fine-Tuning

Manipulation of Human Preference Data

Upstream Contamination of Automated Data Pipelines

Malicious Packages Entering the Development Environment

Backdoored Models on Public Platforms

Exploitation of Deep Learning Framework Vulnerabilities

Supply Chain Vulnerabilities in Preprocessing Tools

Insecure Serialization Format

Systemic Risk from Model Homogeneity

Data Leakage Through Model Memorization

Model Inversion Reconstructing Training Data

Societal Bias Embedded in Historical Data

Insufficient Representation in Training Data

Imbalanced Multilingual Data Distribution

Stereotypes Learned and Reinforced

Copyright Disputes Over Web-Scraped Content

Label Tampering Attack

Gradient Poisoning in Distributed Training

Vulnerabilities in the Training Framework Itself

Tampered Training Records Concealing Problems

Replacement of Fine-Tuning Layer Files

Opaque Model Provenance Chain

Membership Inference Revealing Participation

Reconstruction Risk from Vector Representations

Annotator Bias Transferred into the Dataset

Synthetic Data Amplifying Existing Bias

Secondary Use Beyond Collection Purpose

Implicit Collection and Opaque Consent

Crowdsourced Annotation Poisoning

Violation of Open-Source Model License Terms

Blind Spots from Missing Component Inventory