Exfiltration via AI Inference API
Exfiltration via AI Inference API is when someone uses a company's AI tool to trick it into revealing secrets it learned during its training. Think of it like interrogating a witness who accidentally blurts out confidential information they were supposed to keep private.
This technique involves an adversary leveraging public or private AI model inference endpoints to perform membership inference or model inversion attacks. By systematically querying the API, the attacker extracts sensitive training data, such as PII or proprietary information, or reconstructs the model architecture itself for intellectual property theft.
Adversaries may exfiltrate private information via AI Model Inference API Access (AML.T0040) by exploiting the model's tendency to memorize training data. Through techniques such as Infer Training Data Membership (AML.T0024.000) or Inverting the AI Model (AML.T0024.001), attackers recover sensitive training samples or PII. Furthermore, the API can be abused to Extract the AI Model (AML.T0024.002), facilitating AI Intellectual Property Theft (AML.T0048.004) by reconstructing the model's weights or decision boundaries.