Adversaries can intentionally confuse or “poison” AI systems to cause malfunctions, and there is currently no foolproof defense against these attacks. Computer scientists from the National Institute of Standards and Technology (NIST) have published a document that outlines the vulnerabilities of AI and machine learning and provides guidance on mitigating these attacks. The publication identifies different types of attacks, including evasion, poisoning, privacy, and abuse attacks, and offers approaches to mitigate their effects. However, the authors note that the current defenses against adversarial attacks are incomplete, and developers and users should be cautious of anyone claiming otherwise.
https://www.nist.gov/news-events/news/2024/01/nist-identifies-types-cyberattacks-manipulate-behavior-ai-systems