V8 preview: AI joins the patrol (SY0-801)

Poisoning, theft and leakage

Attacks on the model and its data:

• Data poisoning: tampering with training or feedback data so the model learns wrong or malicious behavior.
• Model inversion: querying a model to reconstruct sensitive training data about people.
• Model theft (extraction): mass queries used to build a copy of a paid model.
• Data leakage: staff paste secrets or customer data into public AI tools that may keep it.
• Shadow AI: unapproved AI tools used without IT review.
• Hallucination: confident but false output (fake citations), dangerous if trusted blindly.

AI as the attacker's tool

Attackers use AI too. AI-generated phishing writes fluent, personalized lures, so spelling mistakes are no longer a reliable sign. Deepfakes clone voices and faces for fake calls from executives. AI also speeds up automated vulnerability discovery and AI-assisted malware development.

Exam tip V8 preview tip: against a cloned CEO voice, the best control is a process: call back on a known number before acting.

Defenses

Treat AI like any powerful new system:

• Least privilege for AI agents: give the bot only the data and actions it needs.
• Human approval before an AI sends money, emails or deletes data.
• Input and output filtering (guardrails) to catch injection attempts and leaked secrets.
• AI acceptable use policy, approved enterprise AI tools and DLP to stop data leakage.
• Verification processes (call-backs, code words) against deepfake requests.