Agent security is shifting from asking whether an AI can be trusted to limiting what it can touch. A vault pattern that brokers credentials and actions is more important than another assistant interface because useful agents will need to operate across accounts, files, networks, and business systems.
The builder-practical lesson is to avoid handing raw secrets to probabilistic workers. Agents should request scoped actions through controlled interfaces, with short-lived permissions, audit logs, approval gates, and revocation paths. This lets the system complete work without turning every prompt injection into a credential exposure event.
In vertical markets, this pattern will separate pilots from production. Healthcare operations, logistics, construction administration, and finance workflows cannot rely on informal trust. They need authority that is narrow, inspectable, and tied to the business process.
Grey Haven read: Useful agents need brokered authority more than broader trust. The security boundary should sit around actions and credentials, not around optimistic beliefs about a model's intentions.
Reader guidance: Treat the item as a signal about operating readiness rather than headline momentum. The useful question is which control, workflow, or evidence requirement must be strengthened before the same pattern reaches a production system with customers, regulators, or internal teams depending on it.
Operator guidance: Replace raw secrets with scoped tools, short-lived credentials, approvals, and revocation. Watch for agents that can read or store tokens directly, especially when they also process untrusted inputs.
Source/provenance: Grey Haven Signal update linear-ASM-92-2, based on the existing backfill draft and source detail Hacker News discussion: https://news.ycombinator.com/item?id=47865822.
