AI Unwrapped! Prompt Injection, Data Poisoning and Deepfakes
7:08Coming soon
An AI assistant reads your inbox and summarises it. One email contains a line you can't see, written not to you but to the assistant, and the assistant does what it says. Nobody hacked the model or the mail server; they sent an email. This lesson shows how AI systems actually get attacked, explained literally, and the defences that work.
You'll learn
- The attack surface of an AI application: instructions, user input, fetched content, training data, output and tools
- Why prompt injection exists: instructions and data arrive as one stream of tokens
- Direct prompt injection (jailbreaks) versus indirect prompt injection, and why tools make it dangerous
- Data exfiltration through links the model writes
- Data poisoning of training data, fine-tuning data and knowledge bases, and backdoors
- AI-powered social engineering: fluent phishing, cloned voices, deepfake video calls
- The defences: least privilege, human approval, untrusted output, separation, filtering, logging, data provenance
- Out-of-band verification against deepfakes
- The real thing: a hidden instruction in a shared document, and the same assistant with read-only tools
- Exam corner: naming an indirect prompt injection
Chapters
- 0:00The email that gave the orders
- 0:41The attack surface of an AI application
- 1:16Why injection works: one stream of tokens
- 2:04Direct and indirect prompt injection
- 3:02Data poisoning
- 3:46AI as the attacker's tool: phishing, voices, video
- 4:31Defences
- 5:32The real thing: a hidden instruction
- 6:07Exam corner
- 6:33Recap and next
Exam objectives: AI-related threats and attack surfaces as they appear in CompTIA Security+ V8. Objective numbers will be added once CompTIA publishes them.
Byte City is not affiliated with or endorsed by CompTIA.