A new AI agent standard could let websites decide which AI agents you're allowed to use
retained text
Meta and Sierra are backing something called the Personal Agent Protocol (Poppy). It's supposed to create a standard for how personal AI agents interact with websites and businesses. Sounds reasonable at first. But I was reading through the [specification](https://personalagentprotocol.org/docs/spec) and some of the details seem pretty concerning from a privacy standpoint. **1\. Websites can block AI agents they don't like** Section 4.1 allows companies to maintain lists of approved and blocked agents. Sure there's a security argument for this. Nobody wants malicious bots accessing their systems. But what happens when someone builds an independent AI agent that helps users find better deals or cancel subscriptions? A company could decide that agent isn't welcome. It's a bit strange that something designed to represent the user could end up needing permission from the very company it's dealing with. **2\. Your AI agent could be handling your actual passwords** Section 4.7 includes an optional sign in process where users give their credentials to an AI agent which then passes them to the company. There are safeguards described in the protocol. Still seems like a weird direction to take when we've spent years telling people not to share their passwords. Giving an AI agent limited account permissions makes much more sense than teaching people to hand over their login details. **3\. Persistent tracking could become part of the system** The protocol supports stable identifiers that allow businesses to recognize returning users even before a normal login. It also supports access that can remain active over time. That's convenient for certain tasks. It's also the sort of thing that could make tracking people across interactions much easier. We've already seen how cookies and advertising IDs turned seemingly harmless features into massive tracking systems. I'd rather see clear limits on what gets stored and how long access lasts before this becomes widely adopted. **4\. The people writing the rules matter too** Meta and Sierra are involved in developing the protocol. The public feedback process even directs people to a Sierra email address. I'm not saying that automatically makes the standard bad. But if this becomes a major part of how people interact with the internet then independent developers and privacy researchers should have a meaningful say in how it works. Especially when the whole point of a personal AI agent is that it should work for the person using it. **The bigger issue for me is who actually has control.** If I choose an AI agent to help manage my accounts or deal with businesses then I should have the final say over what it can do and what information it shares. Companies need ways to protect their systems. That shouldn't automatically mean they get to decide which tools consumers are allowed to use. And any system that makes it easier to identify users across interactions deserves scrutiny before it becomes an internet standard. I'm not suggesting the protocol is intentionally designed for surveillance. There are legitimate reasons for some of these features. But we've seen enough supposedly convenient technology turn into tracking infrastructure to know that privacy protections shouldn't be an afterthought. I'd be interested to hear what people working on open source or locally hosted AI agents think about this. The technical tradeoffs matter but so does making sure users don't lose control of their own tools. Full specification: [https://personalagentprotocol.org/docs/spec](https://personalagentprotocol.org/docs/spec)
document fetched 2026-10-21 00:37 · canonical https://reddit.com/r/privacy/comments/1xb7b0h/a_new_ai_agent_standard_could_let_websites_decide