How to Tame Your Dragon: Productionizing Agentic...
summary
Agentic email assistants are exploitable: a spear-phishing white-text prompt injection tricks rule selection and summarization to authorize an illicit payment. Countermeasure builds an agentic workflow grapher from OpenTelemetry traces, derives behavioral profiles via embeddings, and runs embedding anomaly detection to flag outliers; initial results beat off‑the‑shelf guardrails and catch the attack pattern that simple policy checks miss (guardrails only flag ~1-in-50 cases). Tooling costs are low for embedding inference but storage of detailed traces drives most expense.
AI-written summary · about
chapters (9)
- 0:00Welcome and introduction
- 1:25AI agents expectations vs reality
- 5:17What is an AI agent
- 7:26Security challenges in LLMs
- 20:06Real-world attack on email agent
- 25:46Limitations of guardrails
- 31:08Building an AI supervisor (M)
- 38:14Conclusion and future work
- 39:18Q&A session
related talks
31mAgentic AI Threats: How Multi-Agent Systems Get Exploited
51mPrompt Injection: When Hackers Befriend Your AI
33mAgentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
48mSecuring Agentic AI
51mAgentic AI and Model Context Protocol's Security Vulnerabilities
19mKeynote: Towards Building Safe and Secure Agentic AI
31mTesting, Evaluating and Debugging Generative AI and Agentic Ap...
27mKeynote: Towards Building Safe and Secure Agentic AI