Tool poisoning WebMCP; can Jev catch it?
retained text
I built a little demo a few weeks ago for a hackathon to get to know WebMCP a bit. The demo is not why I’m here though - it’s got me thinking about how useful it can be for site owners to limit and control how their site is consumed in order to provide predictable and desirable outcomes for the end user more efficiently. It also brings the old '*be careful clicking on things you find on the internet*' advice to the forefront again in a kind of supercharged way because the agents decide to trust the tool based on the description not the tool itself. What are we doing to hedge off malicious tool poisoning hiding behind benevolent tool descriptions? Can we manage this risk effectively using zero trust and least privilege without giving up all the agency that made the things useful in the first place? I haven’t played with Jev yet but I think this is the kind of thing it should be really well suited to; checking that the tools description and execution align before the agent gets the tool and trusts it.
document fetched 2026-09-28 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1wrm7d3/tool_poisoning_webmcp_can_jev_catch_it