Research · Updated 2026-07-16
Just-in-Time Access
Standing permissions are standing risk. Issue agents short-lived, task-scoped credentials at the moment of need, and let them evaporate when the work is done.
What it is
Just-in-time (JIT) access replaces durable grants with short-lived credentials issued at runtime and revoked immediately after task completion (Oso). The end state is what Krista Case (theCUBE Research) calls zero standing privilege: an agent holds no persistent access rights at all — only whatever the current, authorized task momentarily requires. Forrester's AEGIS framework prescribes the same posture for its IAM domain: agents are hybrid identities operating on "just-in-time privileges with human oversight."
Why it matters
- Standing credentials are the attack surface. An always-valid token held by an always-on agent can be stolen, replayed, or triggered by a prompt injection at any hour. A credential that lives for minutes shrinks that exposure window to nearly nothing (Oso).
- Agents don't need durable access — they need momentary access, often. Unlike an employee's role, an agent's tasks are discrete and machine-mediated, so runtime issuance is actually practical at scale (theCUBE Research's "real-time control system" framing).
- Expiry is the cheapest revocation. Revoking a compromised standing grant requires detection first; an expired token revokes itself. This complements — not replaces — the instant revocation paths in Delegated Authority.
Best practices
- Issue credentials at task start, revoke at task end — never provision an agent's access at deployment time (Oso).
- Bind the credential to the task's scope, combining JIT with Least Privilege: short-lived and narrow, not short-lived but omnipotent.
- Add behavioral guardrails that revoke early. Oso recommends automatically revoking tokens when an agent's activity pattern turns anomalous — expiry is the ceiling, not the only exit.
- Use predictive permissioning for scheduled needs. Grant time-limited elevation for known events (a reporting window, a migration) rather than permanent elevated roles (Oso).
- Layer context conditions on issuance — device trust, network, time of day — so the same request can be denied in an unexpected context (Oso). Enforcement lives in Runtime Authorization.
Examples
- A data-analysis agent requests warehouse credentials when its job starts; the token is scoped to the three tables in the query plan and expires in fifteen minutes — a stolen copy an hour later is worthless (Oso's JIT model).
- Under zero standing privilege, an idle agent enumerated by an attacker has literally nothing to steal: its directory entry exists, its access does not (theCUBE Research).
- AEGIS-aligned deployments pair the JIT grant with a human approval step for sensitive elevations, so the "just in time" moment is also the oversight moment (Forrester).
Sources
The lessons on this page are summarized from these publications.
- Setting Permissions for AI Agents — Hazal Mestci, Oso Security (2025-10-28)
- AI Agents Are Exposing the Limits of Traditional Identity Governance — Krista Case, theCUBE Research (2026-05-13)
- Introducing AEGIS: The Guardrails CISOs Need for the Agentic Enterprise — Jeff Pollard, Forrester (2025-08-04)
Related topics
- Least Privilege & Least AgencyGrant agents the narrowest capability the task needs — and bound not just what they can access but how much they can decide and act on their own.
- Runtime AuthorizationAuthorization for agents must move from static provisioning to real-time policy decisions enforced at every layer of the interaction, from prompt to response.
- Delegated AuthorityHow a human hands authority to an agent acting on their behalf: consent, scope, and accountability when delegation is the default, not the exception.