How to Build Trustworthy AI
summary
Claims that matter: concrete defenses and tools. Notes include the Chevy Tahoe $1 chatbot glitch, Slack prompt‑injection data leaks, and Fortnite’s Darth Vader NPC showing unsafe outputs. Recommends MLSecOps practices, model scanning with Modelscan, guarding against pickle serialization attacks that run arbitrary code on deserialization, and continuous AI red‑teaming (PyRIT cited). Emphasizes investing in AI runtime security (example: guardrails blocking a patient‑record change) because runtime catches prompt injections and jailbreaks CI/CD misses.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:15Trustworthy AI definition
- 1:04Responsibility in AI
- 2:03Core components of AI
- 4:02Shifting left and right
- 5:57MLSecOps and model scanning
- 8:11Model scanning demonstration
- 9:24AI red teaming
- 11:26AI runtime security
- 13:27Real-world AI architecture
- 15:33Pillar runtime security tool
- 17:10Compliance and trust centers
- 18:56Business risks and regulations
- 20:35The future of AI innovation
- 23:19Conclusion
related talks
18mDefending the unknown using AI and cybersecurity
34mTrust in the Age of AI Agents | Building Trustworthy AI
37mFundamentals of AI safety and security
59mPentesting AI: How to Hunt a Robot
53mHow AI is impacting cybersecurity. Does it?
06m7 practical tips for safer AI and language models
54mA Practical Guide to AI Threats & Mitigation Strategies
38mTrustworthy AI: The Key to Unlocking High-Stakes Applications