Evasion attacks typically involve crafted anomalous inputs designed to manipulate the AI model into producing incorrect or unexpected outputs. Monitoring for anomalous input patterns that result in unusual model behavior provides the most effective real-time indicator of attempted evasion attacks.

For pre-delivery phishing detection and classification, the most appropriate capability is NLP-- tokenization, feature extraction, semantic similarity, and supervised classifiers (e.g., transformer- based classifiers) tuned on phishing corpora and indicators. NLP models score messages and drive automated quarantine policies.

AAISM highlights the importance of an AI threat-control matrix that:
- maps threats (e.g., poisoning, exfiltration, injection)
- identifies required controls
- defines assurance/testing activities
- ensures consistent coverage across all AI components

In AAISM, usage of AI for activities involving personal data and profiling, such as personalized advertising, is explicitly mapped to stringent regulatory and compliance requirements (e.g., data protection, consent, profiling limitations, fairness obligations). The material notes that these activities may trigger "heightened regulatory scrutiny, mandatory impact assessments, and potential penalties for non-compliance." While reputational (B), fraud (A), and commercial (C) risks are all relevant, the primary, non-optional constraint is compliance with applicable regulations governing personal data, automated profiling, and targeted content.
Failure in this area can lead not only to reputational harm but also to legal sanctions, enforced remediation, and operational restrictions. Therefore, regulatory risk is identified as the most important consideration when deploying generative AI for personalized advertising.
References: AI Security Management™ (AAISM) Study Guide - AI, Privacy, and Regulatory Compliance; High-Risk Use Cases and Profiling.

Training on images from a limited local community creates a dataset that is not representative of broader populations, leading to sampling bias where the model performs poorly on underrepresented groups.