Measurement and Audit of Trust, Safety, and Reliability in AI Ecosystems
AI systems are now deployed as part of complex ecosystems rather than as standalone models. Their trustworthiness depends not only on model capabilities but also on the surrounding APIs, tools, platforms, and agent interactions in real-world use. We will introduce recent work on measuring and auditing trust, safety, and reliability in AI ecosystems. We will first discuss shadow APIs, third-party LLM services that claim to provide access to frontier models, and show how they can raise reliability and reproducibility risks. We will then present a measurement study of Moltbook, a social network designed for AI agents, focusing on agent discourse, toxic and manipulative content, and temporal risk patterns. Finally, we will discuss to understand the bidirectional risks between AI agents and the real world.
Date and Time
Location
Hosts
Registration
-
Add Event to Calendar