Eric Siu warns that Claude's tag feature risks locking company knowledge inside the platform. You can't easily export memories if employees leave, face high usage costs, and open yourself to prompt injection attacks. It's renting your core company memory.
Claude tag is actually a Trojan horse. What they're logging you into, it's not just— we're not talking about model lock-in where they're forcing you to use their models, but it's context lock-in. If you start to compound all your knowledge inside of Claude and that it hangs there and like people quit your company, it's not like you can share that memory across, but that entire Claude system has your memory as a company. I think sure, you can export and things like that, but do you want to be paying on usage? And then you start to think about prompt injection, you start to think about other threats that that maybe those people aren't thinking about, once they're locked in, like you're basically renting from them to run your company, right? And you don't want to do that.
