Part II: Data and model attack paths
Data and model attacks differ by what the attacker can reach, what changes or leaks, how long the effect can last, and which local control and evidence apply.
Every branch is compared using the same fields: attacker access, affected object, mechanism, harm, duration or spread, local control, control limits, and evidence. Training and model attacks can alter learned state or accepted model files. Request and response attacks act on inference inputs or outputs while parameters remain fixed; one effect may concern one delivery, while an attack can use repeated queries. Retrieval and stored-context attacks can carry false information or hostile instructions into later requests without changing parameters. Disclosure and extraction attacks expose protected data, inferred facts, or model behavior through transfers, retained copies, and query interfaces. A wrong answer alone does not establish an attack, and an attempted operation does not establish a completed effect.
Chapters in this part
- Training poisoning and model tampering: Altering records used in training can change a model’s learned behavior. Investigating a suspected change requires evidence connecting the affected model version to its training data and training process.
- Input attacks and response tampering: Altering a request or its delivered response can produce a misleading result without changing model weights. Investigating that result requires records of what the caller submitted, what the model received and generated, and what the recipient obtained.
- RAG and search poisoning: Adding false information or hostile instructions to retrieved material can change an application’s answers without changing model weights. Stored copies can carry that influence into later requests.
- Data leakage and privacy attacks: An AI application can expose protected information through the data it sends, the copies it keeps, or the answers its model returns to specially chosen queries. Each route needs its own evidence, because a control on one route does not cover the others.