Erhältlich:
Nicht auf Lager
Buch (Softcover): Fachbuch
10 Real AI Security Incidents
Attack Case Studies and Lessons Learned
Verlag:
Independently Published Unsere-Artikel-Nr.: P37028441
EAN: 9798185810040
Erhältlich:
Nicht auf Lager
Zustellung: Di, 17.11.2026
Versand: Kostenlos
CHF 76.–
Beschreibung
AI security did not fail in theory. It failed in production. If you have ever looked at an AI system and thought, "That should be safe enough," this book exists to gently - and sometimes painfully - prove why safe enough is no longer enough. 10 Real AI Security Incidents. is a practical, story-driven deep dive into how modern AI systems actually break in the real world. Not in lab environments. Not in whitepapers. Not in hypothetical future scenarios. These are real classes of incidents that have already happened across LLM applications, AI agents, retrieval pipelines, and autonomous systems - and they happened because of assumptions most teams are still making today. Inside, you will walk through ten real-world AI security failures in full forensic detail:. Incident 1 - Prompt Injection to Data Leakage. - how a crafted input bypassed output controls and exfiltrated sensitive user data through a channel nobody thought to monitor. Incident 2 - Indirect Prompt Injection via RAG Pipeline. - how attacker-controlled documents embedded in retrieval systems redirected agent behavior without touching the application directly. Incident 3 - Jailbreak Bypassing Policy Controls. - how a multi-turn roleplay sequence dismantled guardrails that passed every pre-deployment safety evaluation. Incident 4 - Cost Exhaustion as Denial of Service. - how API abuse patterns turned legitimate usage into a sustained financial attack that went undetected until the invoice arrived. Incident 5 - Training Data Poisoning. - how corrupted training inputs introduced persistent backdoor behavior that survived fine-tuning and appeared only under specific trigger conditions. Incident 6 - Agent Tool Abuse. - how an AI agent was manipulated into calling legitimate tools with attacker-controlled parameters, escalating access far beyond intended scope. Incident 7 - Memory Poisoning Across Sessions. - how a malicious memory entry persisted across conversations, silently shaping agent behavior for every user who triggered the same retrieval. Incident 8 - Excessive Autonomy and Business Impact. - how an agent operating within its designed permissions made a sequence of decisions that caused real legal, financial, and reputational damage. Incident 9 - Full AI Agent Takeover. - how a multi-stage attack chain compromised agent goals, persisted through memory, and went undetected for weeks before triggering a visible failure. Incident 10 - Model Extraction and IP Theft. - how systematic probing reconstructed proprietary model behavior, exposing intellectual property through the inference API alone. Each incident is broken down step by step - system architecture, entry points, escalation paths, detection failures, timeline, root cause analysis, and the concrete security controls that would have stopped it. No hand-waving. No vague recommendations. Every chapter maps directly to defenses you can implement. 10 Real AI Security Incidents is Book 10 in the series:. The AI Security & Hacking Bible: Protect and Exploit LLMs and Autonomous Agents. Every volume in this series - LLM Security in Practice. , AI Threat Modeling. , The LLM Top 10 Security Guide. , Red Teaming LLMs. , How AI Agents Work. , Hardening AI Agents. , The AI Agent Attacker's Playbook. , Building Bulletproof AI. , and AI Security Operations Guide. - teaches you where the risks live, how to model them, how to test them, and how to defend against them.