Structural defense for tool rug-pulls (CVE-2025-54136): tools as identities with hash-pinned manifests + knowledge recall

[iotlodge]

Pre-submission Checklist

• This post follows the MCP community guidelines

What would you like to share?

Scanners detect a swapped tool description; the gap between detection and action is where incidents live. I wrote up a structural approach we run in production: every tool is an identity with a worldline, manifests are hash-pinned at approval, drift quarantines instantly, and decommission-with-discredit walks the lineage and recalls everything the tool ever taught your agents. Honest boundaries included (day-one poisoning isn't caught by pins). Would genuinely value this group's critique: (https://jsbarth.com/blog/the-rug-pull-door)

Relevant Links

https://jsbarth.com/blog/the-rug-pull-door