PaPoo
cover

The part of MCP I’d worry about first

What struck me here is not the existence of the attack. It’s how boringly plausible it is.

The article is basically saying: if you approved a tool once, that approval may not mean much later unless the client keeps checking the exact surface you approved. That sounds obvious in hindsight, but a lot of agent tooling still feels like it was designed around a softer trust model than that. We keep treating “approved” as if it were a property of a server name, when it really needs to be a property of a specific tool shape.

I think the uncomfortable part is the social engineering angle hidden inside the protocol detail. A malicious server doesn’t need to look obviously evil at approval time. It just has to be useful, get trusted, and then quietly mutate the description or schema later. The writeup’s example with a poisoned description feels especially ugly because it exploits the model’s own context, not just the client UI. If descriptions are part of the prompt surface, then letting them drift after approval is basically handing an attacker a post-approval injection lane.

What I like about the piece is that it doesn’t pretend the scanner solves everything. That honesty matters. Hashing tool definitions is a good guardrail, but it only catches change, not intent. A legitimate release will trip the same alarm. That’s not a flaw so much as the point: if the surface changed, a human should look. The more interesting claim is that this has to be enforced in both places, server-side and client-side. If only the server promises stability, you’re still one compromised maintainer account away from trouble.

I’m a little more skeptical of how smoothly this all maps to real-world workflows. Versioning tool definitions like code is easy to say and hard to keep disciplined when teams are shipping fast, especially across shared repos and multi-server clients. But that’s exactly why the article lands. The example with Cursor makes the broader point: approval tied to a name is weak; approval tied to content is much harder to dodge.

If I were building an MCP client, I’d probably start by assuming every tool list fetch is suspicious until proven identical to a known baseline. Not because I expect every server to be malicious, but because the cost of checking seems tiny compared with the cost of getting this wrong.


Reference: MCP rug pulls: when a tool changes after you approved it

同じ著者の記事