Securing AI at Scale: Practical Defenses against Prompt...
summary
Cloud-scale defenses include deployable middleware that runs semantic filtering (embedding-based intent checks) at API gateways, sandboxed runtimes for untrusted plugins, and per-request policy enforcement tied into CI/CD. Practical hardening uses a mix of adversarial training (FGSM for breadth, PGD for strength), diffusion-based input purification (LURID) and randomized smoothing; LURID runs as a module so no retrain is needed. Proven ops: hash every transformation for data provenance, use canary records in model artifacts, federated anomaly detection, and automated quarantine/rollback for compromised endpoints.
AI-written summary · about
chapters (9)
- 0:00AI security at scale
- 2:27Together’s AI infrastructure
- 8:06AI attack examples
- 13:00Mitigating prompt injection
- 17:36Defending against adversarial AI
- 26:52Preventing model poisoning
- 31:29AI security blueprint
- 35:17Implementing defenses
- 37:45Key takeaways
related talks
09mSecuring Generative AI Workloads
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
28mSecuring AI by design
31mSecurity Considerations for Services Using AI Models
45mSecuring AI: There Is No Try, Only Do!
40mSecuring Generative AI: Is it all an Illusion?
12mSecuring AI-Driven Finance