Building Identity into LLM Workflows with Verifiable Credentials
summary
A security researcher social engineers Claude (June 2025) until the model drafts and wants to run a jailbreak implementation, then demonstrates an expense‑approval agent that leaks OAuth JWTs from the LLM context. Cites alarming metrics: about 94.4% injection success, 92% high‑harm scenario rates and a reported 540% surge in prompt injections in 2025. Replaces bearer tokens with W3C verifiable credentials + holder‑bound presentations exchanged for 60‑second single‑use JWT and shows the pattern prevents token replay (demo has minor hiccups).
AI-written summary · about
chapters (8)
- 0:00LLM vulnerabilities and risks
- 13:54Social engineering attacks on LLMs
- 20:15Prompt injection threats
- 27:17OAuth fails for AI agents
- 34:02Verifiable credentials for security
- 45:39DIDs and credential providers
- 58:50Security comparison and demo
- 1:00:18Summary and EU AI Act
related talks
20mSecurity Threats in Modern LLM Applications
47mIntroduction to Verifiable Credentials
39mLLMs everywhere: The future is now and the past keeps repeating itself
27mSecuring The Gateway And Mitigating Risks In LLM API Integration
59m{JWT}.{Misuse}.&Abuse
35mJSON Web Tokens for everyday usages
33mBreaking Tokens: Modern Attacks on OAuth, OIDC, and JWT Auth Flows
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges