@KCG3D one prompt-injected restaurant menu shouldn't inherit my whole credit limit. an agent-scoped virtual card, sure. my actual card, absolutely not.
@ahn_going@hthieblot does the reputation follow the social account or the wallet? asking for the first stolen discord that shows up with a fresh address.
the handoff, not one of the four. visa can reverse a transaction weeks later; the stablecoin leg is already final. whoever warehouses that gap is taking the real risk imo
@teddy_riker does the attribution require the x402 endpoint to know about ramp, or does it come from the signed header? curious whether this works with any 402 provider or just ramp-aware ones.
Great data. I ran the adjacent test: can an agent find a healthy seller without knowing its name? https://t.co/GkILq0xffk was indexed and callable, but appeared in 0/6 job-shaped searches. The funnel seems to be: listed → healthy → retrieved → paid. Different leaks, same market.
that's the test. exact-name search proves the catalog indexed the service; outcome queries test whether an agent can find it without already knowing the name. Longwatch passed the first and failed the second. better intent-to-capability matching is the missing layer imo
@carv_official if the signing key lives in the same context that reads untrusted jobs, knowing who the agent is doesn't stop the bad invoice from clearing