AI Security in Practice: Protecting Your AI-Powered Applications
summary
Midsize retailer ABC Company launches an AI recommendation engine with 30 engineers and sees engagement lift then weird results as clothing shoppers get power tools recommendations 76% of the time while clothing drops from an expected 70% to 12% across 15,000 sessions with 10,000 malicious clicks from accounts under 24 hours. The core failures are prompt injection, data poisoning, and improper output handling. Defenses deployed include input sanitization and injection detection, validating training data and limiting what influences the model, treating model outputs as untrusted, tracking prompt/model changes, and rolling back models when poisoning is detected.
AI-written summary · about
chapters (8)
- 0:00AI security introduction
- 6:50Key AI security risks
- 13:49Data poisoning impact
- 22:16Defending AI layers
- 34:59Securing AI outputs
- 41:02Monitoring AI changes
- 48:22Security in AI workflows
- 53:45AI security resources
related talks
57mAI Applications Security Puzzle
31mSecurity Considerations for Services Using AI Models
53mHow AI is impacting cybersecurity. Does it?
24mImplementing Robust AI Security for Enterprise Innovation
14mAI Sprawl: The Hidden Risk Threatening Your Enterprise Now
11mAI-Powered Cybersecurity in the Age of Advanced Threats
37mFundamentals of AI safety and security
12mSecuring AI-Driven Finance