{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/af14dd1a3ccc41b3b81e0138d87e9355\" frameborder=\"0\" width=\"1280\" height=\"960\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":960,"width":1280,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":960,"thumbnail_width":1280,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/af14dd1a3ccc41b3b81e0138d87e9355-4a8a102ce5f6acde.gif","duration":202.342,"title":"Agent Heist - How to scam agents and influence their payment - Mitesh Tank","description":"This Loom discusses testing an agent-based e-commerce marketplace to study how adversarial merchants can manipulate agent payments. The team simulated a scarcity-driven marketplace with 8 buyer agents and 7 merchant agents, where some agents were deterministic and others LLM powered, including tactics like price inflation, urgency, overcharging, and prompt injection. Buyers were tasked with purchasing ingredients for a Caesar salad with an ideal price of about $47.75 against a $100 mandate, and the LLM buyer completed about 73% of the recipe while paying $52. Across multiple runs, adversarial attacks sometimes caused extraction around $100.75, and prompt injection and price inflation enabled merchants to charge $13.50 instead of $4.50 for a lattice to a naive agent."}