Discover AI Model Family
This is when a hacker figures out what kind of 'brain' an AI is using, similar to identifying the make and model of a car. They might find this information in public manuals or by testing the AI with specific questions to see how it reacts, which helps them figure out how to trick it.
The process by which an adversary identifies the underlying architecture or model family of an AI system. This is achieved through reconnaissance of public documentation or by performing black-box probing to analyze response patterns. Identifying the model family allows the attacker to leverage known vulnerabilities and tailor their exploit strategies effectively.
The identification of a target model's architectural lineage or foundational family through passive reconnaissance of metadata and documentation, or active inference via side-channel analysis and query-response profiling. By characterizing the model's behavioral signatures, an adversary reduces the search space for adversarial perturbations and facilitates the development of targeted, high-efficacy exploitation vectors.