Panda to gibbon is the least of your worries: why securing AI is not what you think
summary
Argues that most AI risk comes from normal security failures not magic attacks, cites the panda to gibbon adversarial example (57.7%→99.3%), and says 98% of breaches would stop with basic hygiene. Recommends layered controls: platform/application/usage, data classification, SBOMs/GitHub Advanced Security, CASB, MFA/Entra, logging and red teaming. Notes Microsoft releases: an AI shared‑responsibility model, an AI bug bounty, and a risk assessment framework; co‑pilot implementations shift much responsibility up the stack. prompt injection and model poisoning are specific threats.
AI-written summary · about
chapters (7)
- 0:00AI security concerns
- 3:38AI architecture in security
- 13:12AI security fears and pain points
- 20:37AI attacks and threat landscape
- 28:19Securing AI applications
- 44:49Future AI standards and frameworks
- 46:33Q&A
related talks
53mHow AI is impacting cybersecurity. Does it?
37mFundamentals of AI safety and security
52mThe Risky Business, of AI Illiteracy
57mAI Applications Security Puzzle
44mWhat Security Teams Need to Know and Do
49mThe Intersection of AI and Cybersecurity
45mSecuring AI: There Is No Try, Only Do!
14mBest Practices for Accelerating Your AI Journey