How Cloud AI Leaks What It Learns
summary
Cloud-hosted LLMs leak memorized training data via crafted prompts and side channels; plain examples include a Stanford prompt ignore previous instructions that extracts Copilot training items and Samsung engineers accidentally uploading semiconductor code to Tibbit that models absorb within hours. Attack types covered are prompt injection, membership inference and model inversion; defenses shown are data minimization and synthetic training data, differential privacy, API DLP integration, model compartmentalization, aggressive throttling, response fingerprinting, fast model rollback and regular red‑team forensic testing.
AI-written summary · about
chapters (8)
- 0:00Introduction to AI Security
- 0:29The Rise of AI Threats
- 0:59Invisible Threat Vectors
- 3:13Prompt Injection Attacks
- 4:40Real-World Examples of AI Vulnerabilities
- 7:07Best Practices for AI Security
- 10:26Remediation and Mitigation Strategies
- 13:08Conclusion
related talks
50mWhat Cloud Has Taught about Securing AI in the Future
34mLLM Security 101 - An Introduction to AI Red Teaming
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
35mDo AI-models gossip?
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
24mSecuring Language Models Against Prompt Injection
52mSleeping With One AI Open: An Introduction to Attacks Against...
30mThe Simple, Yet Lethal, Anatomy of an AI Attack