Fundamentals of AI safety and security
summary
Microsoft outlines concrete risks and mitigations: AI expands attack surfaces (example Morris worm jailbreak, a $25M deepfake payment, 2010 flash crash) and requires layered defenses using a “Swiss cheese” model. A demo architecture, Contoso Draw, shows untrusted prompts, plugin image-to-prompt exploits and cross‑prompt injection leading to account takeover. Practical defenses include data delimiting data marking and data encoding, plus red teaming, real‑time monitoring (Defender for Cloud AI Security alerts) and early threat modeling to reduce costly fixes later.
AI-written summary · about
chapters (7)
- 0:00AI safety fundamentals
- 1:05Secure AI development
- 6:19Unique AI risks
- 9:24Categorizing AI threats
- 20:51Early risk identification
- 25:17Testing and red teaming
- 32:44Deployment and monitoring
related talks
49mThe Intersection of AI and Cybersecurity
24mHow to Build Trustworthy AI
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
53mHow AI is impacting cybersecurity. Does it?
57mAI Applications Security Puzzle
44mWhat Security Teams Need to Know and Do
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
31mSecuring the Future: Navigating AI Vulnerabilities and Evolving Security Practices