At Identiverse 2026 last week, GitGuardian's Andrej described a customer environment with 70,000 non-human identities, most without a named owner. That number came up because it's the actual scale problem underneath every AI-agent security pitch this year: organizations don't have too few controls, they have too many identities nobody claimed first. The instinct is to fix this with better discovery — scan the environment, find the shadow agents, build an inventory. But a competing view surfaced at the conference that's more interesting: govern at credential creation, not after discovery. Instead of waiting to find orphaned agents and retrofit ownership onto them, tie every credential issuance to an owner and an attestation obligation at the moment it's minted. Rotation programs, in this view, don't solve the underlying problem — they just delay it. A credential that starts anonymous stays exploitable no matter how often you rotate its keys. The mechanism GitGuardian's research points to is cryptographically attested, short-lived, scope-bound tokens issued at runtime, rather than static API keys handed to an agent once and reused across every sub-agent it spawns. The failure mode is specific: an agent bootstrapped with a long-lived key carries that key's full permission set into sub-agents its original developer never anticipated. The blast radius grows every time the agent delegates, and no rotation schedule catches that because rotation just refreshes the same over-scoped grant. This is a narrower claim than "we need better agent discovery," and it's worth separating from that broader theme because the two solve different failures. Discovery tells you an agent exists. Credential-time governance determines whether that agent could ever have been anonymous in the first place. Several conference writeups converged on the same three-part framework — inventory, governance, runtime authorization — but the credential-creation argument specifically inverts the usual order: don't inventory first and govern later, because by the time you inventory a 70,000-identity environment, the ownership gap is already baked into every one of those credentials. Forrester's Identiverse recap reached a related but distinct conclusion: AI agents don't fit the existing mold of static, human-timescale identity governance tooling, and the recommended design pattern is delegation to a uniquely identified agent rather than impersonation of a human's existing access. Put together, the emerging consensus isn't a single fix — it's an admission that identity infrastructure built for humans logging into systems doesn't have a slot for an agent that mints its own sub-agents mid-task, each inheriting credentials nobody explicitly issued to them.