Every company you've ever worked for has an offboarding checklist. Badge surrendered. Laptop wiped. Accounts disabled by end of day. HR runs it, IT executes it, and somewhere a compliance officer sleeps better because of it.
Now count the AI agents your organization has deployed in the last eighteen months. The prototype that scraped invoices. The support triage bot from the Q3 push. The three "temporary" workflow agents somebody's team stood up before the reorg. Which checklist ran when those stopped mattering?
None. Because agents don't resign. They don't file exit paperwork. They don't self-report dormancy. They just go quiet — while their API keys, OAuth grants, and service-account permissions stay exactly as live as the day you issued them.
The workforce that never leaves
The scale of this problem is no longer hypothetical. The Cloud Security Alliance's 2026 State of Non-Human Identity and AI Security survey found that non-human identities already outnumber humans 45 to 1 in the average enterprise — and only 12% of organizations have automated lifecycle management for any of them. The other 88% are doing it by hand, which in practice means not doing it at all.
The credential half-life data is worse. GitGuardian's State of Secrets Sprawl 2026 retested secrets it had confirmed valid back in 2022 and found 64% were still not revoked by January 2026. Four years. Meanwhile, the same report documents attackers exploiting exposed AWS credentials within an average of 17 minutes. Your revocation process runs on a quarterly review cycle; your adversary runs on a coffee break.
And when the Non-Human Identity Management Group looked specifically at offboarding practice, only 20% of organizations maintained a formal revocation process for API keys at all — and 91.6% of leaked secrets were still valid five days after the owner was notified.
Read those numbers together and the picture is plain: we have built a workforce that gets hired with great ceremony and never, ever gets fired.
Why the human playbook fails
It's tempting to bolt agents onto the employee offboarding process and call it governance. That fails for three structural reasons.
There is no termination event. Human offboarding hinges on a moment: the resignation letter, the last day. Agents fade. Usage drops, the sponsoring project ends, the engineer who deployed it changes teams. No trigger fires, so no process starts. The agent isn't terminated — it's abandoned with privileges.
Ownership evaporates. An employee has a manager. An agent has whoever set it up, until they leave, after which it has nobody. When responsibility isn't attached at issuance, offboarding becomes a shared assumption — and the last revoke never happens.
Credentials outlive context. The agent's key doesn't know the project ended. Long-lived static credentials are, by design, context-free: they authorize forever, everywhere they're accepted, regardless of whether anyone still intends the authorization.
Offboarding is a lifecycle problem, not a cleanup problem
The fix isn't a better quarterly audit. Audits find orphans after months of exposure; the goal is to make orphans structurally impossible. That means treating decommissioning as a first-class lifecycle state — designed in at deployment, not improvised at discovery. Concretely, real agent offboarding has four properties:
- Every agent has an owner and a state. Not a wiki page — a registry, where "retired" is a machine-enforced status, not a comment.
- Credentials expire by default. A credential that must be renewed is a credential that dies when everyone stops caring. Revocation shouldn't be an emergency procedure; expiry should be the resting state.
- Dormancy is a signal. An agent that hasn't acted in 30 days and suddenly starts moving is not a returning colleague. It's an incident.
- Termination produces evidence. "We turned it off" is a claim. Auditors — and increasingly regulators — want proof of when access ended and what happened after.
How VeriSwarm closes the loop
This is the lifecycle problem Fleet, Passport, and Vault were built to close.
Fleet makes agents managed instances instead of scripts with keys. Every agent deployed through Fleet carries an owner and an explicit lifecycle state — and retirement is an operator workflow, not a hunt through IAM consoles. Decommission an agent and its runtime, schedule, and access are shut down together, in one action.
Passport replaces context-free credentials with signed manifests: ES256 JWT credentials that carry expiry, scope, and delegation chains. A Passport credential dies on schedule unless deliberately renewed — so an abandoned agent's authority drains away on its own, instead of persisting until someone remembers it exists. And because delegations are explicit, revoking a parent cuts off every agent acting downstream of it.
Gate watches for the fade. Trust scores decay with dormancy, and an idle agent that resumes activity gets scored — and policy-gated — like the anomaly it is, not waved through on a two-year-old approval.
Vault turns the shutdown into evidence. Every lifecycle transition and every post-retirement access attempt lands in the immutable, hash-chained audit ledger. When the questionnaire asks how you ensure decommissioned systems no longer access production data, you don't describe your intentions. You export the chain.
The exit interview your agents never had
The industry has spent two years perfecting agent onboarding — templates, frameworks, one-click deploys. Nobody shipped the leaver process. That asymmetry is now the gap attackers live in: not the agent you're watching, but the forty-five identities per employee that nobody is.
Your agents deserve what every departing employee gets: a clean exit, on a known date, with the keys handed back and a record that it happened.
Give them one before an attacker conducts the exit interview for you. Start with Gate free, and retire your first orphaned agent this week.
Sources: Cloud Security Alliance / Security Boulevard — non-human identities outnumber humans 45:1; 12% automated lifecycle management · GitGuardian State of Secrets Sprawl 2026 — 64% of 2022-valid secrets unrevoked; 17-minute exploitation window · NHIMG — 20% formal API-key revocation; 91.6% of secrets valid five days post-notification