STOIC Security: Shielding Your Generative AI App from the Five Deadly Risks
summary
Introduces the STOIC framework — Stolen, Tricked, Obstructed, Infected, Compromised — and cites concrete attacks: model extraction via millions of API queries that clone GPT-3.5 behavior, model inversion that reconstructs training images, prompt‑injection hidden inside images that makes agents execute commands, and reports of Hugging Face models containing reverse‑shell payloads. Recommends practical defenses: input/output gating, canary prompts, rate‑limits, cryptographic signing of datasets/models, zero‑trust agent scopes, model segregation, continuous monitoring and cross‑functional governance.
AI-written summary · about
chapters (12)
- 0:00STOIC Security framework
- 5:39Generative AI security challenges
- 17:10Real-world AI attack harms
- 22:47STOIC framework for LLM threats
- 23:28Stolen attacks and model risks
- 27:57Tricked attacks and prompt injection
- 32:48Obstructed attacks and DoS risks
- 36:38Infected attacks and model poisoning
- 41:09Compromised infrastructure risks
- 43:59AI security mitigation strategies
- 50:40Key lessons for AI security
- 53:28Q&A
related talks
25mThreat Modeling for AI Apps with Attacks as Code
09mSecuring Generative AI Workloads
53mHow to Break AI Systems (Before Someone Else Does)
30mBattling the Machine: Cutting-Edge Attacks and Defenses in Generative AI
21mProtect your generative AI applications against jailbreaks
40mSecuring Generative AI: Is it all an Illusion?
19mKeynote: Towards Building Safe and Secure Agentic AI
18mDefending the unknown using AI and cybersecurity