What does NIST AI 100-2e2025 (Adversarial ML) say about AI in banking?

What are the key points of NIST AI 100-2e2025 (Adversarial ML)?

• Authored by NIST's Computer Security Division with Northeastern University, Cisco, the UK AI Security Institute and the US AI Safety Institute.
• Predictive AI taxonomy: evasion, poisoning (data and model) and privacy attacks (membership inference, data reconstruction, model extraction), with mitigations for each.
• Generative AI taxonomy: AI supply-chain attacks, direct prompting attacks including jailbreaks, indirect prompt injection via retrieved or tool content, and misuse enablement; the 2025 edition adds agent- and RAG-related attack vectors.
• Classifies attacks by learning stage (training vs. deployment), attacker knowledge (white-box, gray-box, black-box) and objective (availability, integrity, privacy, misuse).
• Is a reference document, not a control standard; NIST positions it to underpin later security guidance such as the COSAiS control overlays and the Cyber AI Profile.

What did NIST AI 100-2e2025 (Adversarial ML) change for banks?

The taxonomy gives bank security and model-validation teams a shared, citable vocabulary for threats to fraud models, credit models and LLM applications — prompt injection, data poisoning and model extraction — that existing model-risk guidance never named. It is the reference most bank threat models for AI now cite, and the basis for the AI-security work NIST is building on it.