Trustworthy Generative AI: The Role of Observability and Guardrails.
summary
Argues that enterprise GenAI needs runtime AI guardrail as API gateway to filter malicious prompts and outputs, plus end-to-end observability that traces prompts, tool calls and vector DB hits and measures metrics like hallucination rate and unusual tool use. Recommends adversarial testing in the DevOps pipeline, encrypting sensitive stores and encrypt vector databases, strict role-based access controls, and human verification gates for high-risk outputs. Claims these combined controls reduce prompt injection, data exfiltration and repudiation while enabling faster root-cause attribution.
AI-written summary · about
chapters (9)
- 0:00Introduction to trustworthy generative AI
- 1:31GenAI trust challenges
- 6:12Under the hood of generative AI applications
- 12:01GenAI threat categories
- 13:50Mitigating GenAI risks
- 15:57Observability for generative AI systems
- 19:49Adversarial testing in GenAI
- 21:12Traditional security practices for generative AI
- 22:16References & closing
related talks
43mAI Engineer Ethics: Shaping Responsible Generative AI
28mWhere are we going with observability?
45mOverview of Risks, Vulnerabilities and Paths to Safer Use of Generative AI
24mImplementing Robust AI Security for Enterprise Innovation
46mCan you trust your AI?
39mSecuring AI at Scale: Practical Defenses against Prompt...
29mElevating generative AI and LLM observability
25mWhen Your "New Teammate" Hallucinates: Observability as the Safety Net for AI Code