Phishing LLMs: Reeling in the Machine
summary
Generative apps share a classic three-tier runtime: API gateway, orchestration (plugins like LangChain and LlamaIndex), and data (often vector DBs), so familiar AppSec controls work. Prompt injection succeeds because control and data are both text; mitigation shown includes egress firewalls, rate limiting and limiting the 8K context window. MITRE lists ~80 techniques and OWASP a top 10. A live test tries roughly 23 attempts on a flight and fails under crude rate limits, showing limits materially reduce attacker ROI.
AI-written summary · about
chapters (24)
- 0:00Phishing LLM applications
- 1:25Gen AI vs traditional systems
- 3:59LLM attack targets: data
- 6:09Control vs data plane security
- 7:14Prompt injection risks
- 8:16Phishing vs prompt injection
- 12:07AI ML deep learning LLM
- 13:31LLMs process text to embeddings
- 16:44LLMs as sentence completers
- 17:26WAFs ineffective vs prompt injection
- 18:51Dual LLM filtering methods
- 19:12Egress firewalls for LLM outputs
- 20:13Rate limiting for LLMs
- 21:21Prompt size limits
- 22:23ChatML role assignment security
- 23:46LLM plugin agency risks
- 25:12Data poisoning in training
- 26:36Knowledge base governance
- 28:04Microsoft threat model for LLMs
- 30:10Adversaries in Gen AI runtime
- 31:18Governance for AI tech
- 32:01Model scanners for Gen AI
- 33:04LLMs as autonomous agents
- 35:12Vendor assessment for AI services
related talks
49mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
27mSecuring The Gateway And Mitigating Risks In LLM API Integration
20mUnderstanding and Mitigating Risks in LLM-Powered Web Apps
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
20mSecurity Threats in Modern LLM Applications
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
51mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges