Attack Surface Area | Gen AI Slides

Prompt injection is a type of attack where malicious users manipulate the input prompts given to a generative AI model. By crafting specific inputs, attackers can cause the model to generate harmful or unintended outputs. This can lead to the dissemination of false information, offensive content, or even the exposure of sensitive data.

Data poison attacks involve the introduction of malicious data into the training dataset of a generative AI model. This corrupted data can skew the model's learning process, leading to biased or incorrect outputs. The goal of such attacks is to degrade the performance of the AI system or to make it behave in a way that benefits the attacker.

Data sources attacks target the integrity and reliability of the data sources used by generative AI models. By compromising these sources, attackers can inject false or misleading information into the model's training data. This can result in the model generating outputs based on incorrect or manipulated data, thereby undermining its trustworthiness and accuracy.

Attacks on the model itself involve exploiting vulnerabilities in the AI model's architecture or algorithms. These attacks can take various forms, such as adversarial attacks where inputs are crafted to deceive the model, or model extraction attacks where attackers attempt to replicate the model by querying it extensively. The objective is to either degrade the model's performance or to steal its intellectual property.