A comprehensive survey posted on arXiv synthesizes the growing field of transferable attacks on AI systems. These are adversarial manipulations or learned attack mechanisms that are constructed in one setting and then successfully compromise other AI systems. The core idea is that an attacker does not need to start from scratch for each target; a single attack recipe can travel.

The survey highlights why this matters: transferability reduces the cost of launching attacks, since the same effort can be reused against multiple systems. At the same time, it amplifies security and privacy risks, because a vulnerability discovered in one model or platform may expose a much wider range of AI deployments.

By bringing together existing research on how and why attacks transfer, the survey provides a structured view of a risk that is often discussed only in isolated examples. It points to the need for defenses that account not just for individual models, but for the broader ecosystem in which AI systems are deployed and shared.