Suhas, Cullen —

Thanks for posting this. It lands on a layer I've been working the
other side of, so a note on how the two fit, one data point, and one
question.

Your introduction's failure mode — a server changing tool descriptions
after a client has approved them — is measurable in the wild. I run a
daily scan of the public MCP ecosystem: in the Nov 1 scan (54,513
servers), we counted 1,509 listing swaps — servers changing what they
serve under an identity clients had already approved. Your §8.2.5
consistency rules are the first client-side mechanism I've seen aimed
squarely at it, for servers that adopt the card.

The passport draft I posted the same week (draft-le-agent-passport-00)
works the call-time side: a passport binding an agent's key to its
accountable controller and a digest of its build, plus delegation
receipts a relying party verifies step by step. SD-Card reads to me as
the discovery-time complement — what an agent discloses, selectively,
before anyone calls it — and I plan to cite it that way in -01.

One question: the draft has no status or revocation mechanism, and
cards can be long-lived. If a server's key is compromised or its
metadata is withdrawn, expiry looks like the only kill switch. Is that
a deliberate scoping choice for -00, or is a status mechanism planned?

Happy to compare notes at 127.

— Ben Le
Redthread Security
_______________________________________________
OAuth mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to