AI Is Just Software, What Could Possibly Go Wrong?
summary
Claim: treat generative models as software with software risks — not magic. Gives concrete failures: package hallucinations with fake npm packages (30k then 300k downloads), code generation creating duplicate primary keys cost $10,000 and five days to debug, prompt‑injection in application letters, data‑poisoning that flips missile classifiers, and cloud AI marketed as “private” but exposing prompts to vendor. Also shows mitigation work: AI Exchange — 150 pages from 50 experts, 78 pages adopted into standards. Focus on where it broke and what fixed it.
AI-written summary · about
chapters (19)
- 0:00AI's dual role: solution and risk
- 2:35AI as software: security practices
- 4:20AI hallucinations and dangers
- 6:27How ML models make mistakes
- 11:47Wrong models and data poisoning
- 15:20Countering data poisoning attacks
- 17:27EU AI Act cybersecurity requirements
- 21:18AI's creative potential and limits
- 22:22Malicious AI and prompt injection
- 26:18AI's impact on workflows
- 27:25AI-driven dev quality challenges
- 30:55Ethical and legal AI pitfalls
- 33:00AI-generated code flaws and review
- 36:52Code review atrophy risks
- 39:04Cloud AI security and transparency
- 42:19AI terminology and security aspects
- 48:43AI system threats and defenses
- 50:49Security role in safeguarding AI
- 51:09Q&A
related talks
14mAI Sprawl: The Hidden Risk Threatening Your Enterprise Now
31mHacking AI Applications
53mHow AI is impacting cybersecurity. Does it?
55mAI Showdown: The Good, The Glitchy, and The Risky
18mAI Is Coming For Our Secrets
17mGenerative AI: Keeping it real
57mAI Applications Security Puzzle
28mAll Your AI are Belong to Us