How It Really Happened: AI Accomplice
AI Accomplice
Security Topic: Invisible Prompt Injection Attacks
NINJIO Season 11: Episode 1
Emotional Susceptibility: Urgency
In this episode, viewers learn about invisible prompt injection attacks as Lily Delamere returns as RoxyProxy, still on the run from the Cybercrimes Taskforce. Based on the cyberattack suffered by Jaguar Land Rover, your colleagues will learn how to protect themselves and their organizations from attacks that manipulate A.I. programs.
Teachable Takeaways
- These attacks often exploit a user’s sense of urgency. Stop and verify suspicious instructions with a human through a separate channel.
- Invisible prompt injections can trick AI assistants into giving people malicious instructions to reset passwords, enter credentials, or click dangerous links.
- Avoid granting AI assistants permission to act autonomously, including email forwarding, calendar creation, or file system operation, unless approved by your organization.
Additional Reading
- Invisible Prompt Injection: A Threat to AI Security – Trendmicro
- When AI goes off-script: Understanding the rise of prompt injection attacks– CyberRisk Alliance
- UK NCSC Raises Alarms Over Prompt Injection Attacks – Infosecurity Magazine
- LLM01:2025 Prompt Injection– OWASP
- Prompt Injection: Understanding and Mitigating Risks – AI Prompt Theory
- Invisible commands, real threats: The rise of prompt injection in AI – AXIOS
- LLM Prompt Injection Prevention Cheat Sheet – OWASP Cheat Sheet Series
- Not what you’ve signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection – Cornell University
About NINJIO
NINJIO reduces human-based cybersecurity risk through engaging training, personalized testing, and insightful reporting. Our multi-pronged approach to training focuses on the latest attack vectors to build employee knowledge and the behavioral science behind human engineering to sharpen users’ intuition. The proprietary NINJIO Risk Algorithm™ identifies users’ social engineering vulnerabilities based on NINJIO Phish3D phishing simulation data and informs content delivery to provide a personalized experience that changes individual behavior.
