MCP security risk: Agent error - OAuth Tutorial

- That brings us to agent error. We often talk about AI hallucinations. When you're working with AI agents, a lot of those hallucinations happen without us seeing it. If you've ever worked with, say, ChatGPT or Claude, in one of these more advanced like research modes, you often see the system is like doing a bunch of things in a separate box, and then that box collapses, and you just sit and wait, and it takes a while, and it's doing a bunch of things. If you go open those boxes, you'll often see the agent go and do a bunch of things, and then after a while, abandoning whatever it was doing because it was wrong, and then do a bunch of other things, and then maybe abandoning them, and then going through a third round before it actually gets where it wants to go. Those are errors, and they are obfuscated from our view because if we actually saw them happening, we would trust the systems less than it would be against the interests of the people who build these systems for us. The…