AI Threats and Attacks

View at publisher Summarise with VIRAT
Authors: Mike Chapple, Fred Nwanganga
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This chapter analyzes security threats and attacks targeting AI systems, focusing on prompt injection, data poisoning, and model poisoning. It explores how attackers manipulate inputs, tamper with training data, or bypass safeguards to cause harm, often through subtle exploitation rather than traditional hacking methods. The chapter discusses frameworks for AI threat modeling and compensating controls to defend against these attacks. Prompt injection, where attackers craft inputs to manipulate AI models, is analyzed in both direct and indirect forms. Data poisoning involves manipulating training data to influence model behavior, whereas model poisoning directly modifies model parameters. The chapter also covers how attackers bypass AI guardrails, which are technical and managerial controls designed to limit misuse and harmful outputs. Key findings highlight the importance of understanding these threats and implementing robust security measures to build trustworthy AI environments. The chapter examines various methods by which attackers may bypass or manipulate AI systems, focusing on guardrail circumvention, hallucinations, input manipulation, and bias introduction. Jailbreaking, a type of guardrail circumvention, involves users crafting prompts to override safety controls, often by embedding requests in layered scenarios or chaining multiple prompts.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Security and Verification in Computing Ethics and Social Impacts of AI

Publication details

Journal / source
Wiley eBooks
Pages
57–86
Publication date
24 February 2026
Language
EN
Document type
Book chapter