Mitigate OWASP Top 10 for LLM risks with a Zero Trust approach
summary
A patient‑info assistant prototype uses Amazon Bedrock agents, API Gateway, Lambda, DynamoDB, KMS, PrivateLink, VPC Lattice, Verified Access and GuardDuty but initial role tags in prompts are bypassed by prompt injection so the LLM leaks records. Replacing in‑prompt auth with immutable session attributes passed to the agent and enforcing access with Verified Permissions in the Lambda removes excessive agency and prevents unauthorized data retrieval. The zero trust pattern proves effective: traditional IAM, least‑privilege service roles and data‑layer authorization stop the disclosure.
AI-written summary · about
chapters (5)
- 0:00Understanding LLM threats
- 5:42Patient information scenario
- 7:51Prompt engineering limitations
- 11:36Implementing zero trust
- 18:08Final takeaways and defense
related talks
30mSecuring AI Workloads: Building Zero-Trust Architecture for LLM Applications
41mJourneys to Zero Trust on AWS
20mSecurity Threats in Modern LLM Applications
55mSteps toward a Zero Trust architecture on AWS
59mAchieving Zero Trust with AWS application networking
23mZero-Trust Architecture for Agentic AI: Securing Multi-User Access and Third-Party Integrations
17mZero Trust access with zero waiting, zero pain, and zero compromises
27mLeveraging Infrastructure as Code to Create Zero Trust Access to Your Apps and Infrastructure