Security Considerations for Services Using AI Models
summary
Services need concrete defenses: prompt injection and many-shot jailbreaking can raise harmful-output probability to 60–80% after ~256 Q&A examples; image noise and stickered stop-sign attacks show data poisoning causes real-world misclassification; model theft and reverse-engineering (e.g., LLaMA leaks) and latency-inducing sponge inputs are observed threats. Recommended controls include AI SBOM/SPDX 3.0 model cards, sandboxed training with access control, rate limiting, monitoring outputs, and tooling such as ART, GAK, AuditAI and CycloneDX.
AI-written summary · about
chapters (7)
- 0:00Introduction and agenda
- 2:29Basics of AI engineering
- 6:06Four types of AI attacks
- 9:42Mitigating AI supply chain risks
- 13:38Secure AI development lifecycle
- 23:20Tools for securing AI services
- 25:51Key takeaways and closing
related talks
57mAI Applications Security Puzzle
46mLessons Learned from Developing Secure AI Workflows
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
45mSecuring AI: There Is No Try, Only Do!
55mAttacking And Protecting Artificial Intelligence
39mSecuring AI at Scale: Practical Defenses against Prompt...
49mThe Intersection of AI and Cybersecurity
06m7 practical tips for safer AI and language models