Slide 19 of 28
Part 4 — PreventionSlide 19
Slide 19 · Mitigation 1
Task-scoped, time-bound permissions — shrink the window and the scope simultaneously
The highest-impact single mitigation. Addresses un-scoped inheritance, delegation abuse, and stale authorization in one control.
📄 OWASP Top 10 for Agentic Applications 2026 · ASI03
MIT 01
Task-Scoped, Time-Bound Permissions

Issue short-lived credentials per task — not per session, not per agent, not per deployment. Each task gets a token with exactly the permissions that task requires, with an expiry that matches the task's expected duration.

Use per-agent identities with mTLS certificates or scoped OAuth tokens. When a manager agent delegates to a worker agent, the worker receives a derived token with a scope that is a strict subset of the manager's scope — not a copy of it. Cap rights with permission boundaries that cannot be exceeded regardless of what the agent requests.

Directly closes Scenarios 01 (delegated privilege abuse), 02 (memory-based escalation via stale credentials), 05 (workflow authorization drift), and 07 (identity sharing). It limits blast radius across all scenarios by shrinking the window of valid access.

Setting permissions based on what the agent deployment might need rather than what this specific task needs. Over-provisioning "just in case" is the primary driver of un-scoped privilege inheritance. Enforce that sub-agents can never receive a wider scope than the specific task they were delegated.

Time-bound tokens must be actively revoked, not just set to expire. If a workflow is cancelled or a user's authorization changes mid-task, the token must be revoked in real time — expiry alone creates the TOCTOU window that Scenario 05 exploits.

💼 Business takeaway

Ask your team: when your AI agent finishes a task, does its access automatically expire — or does it keep that access indefinitely? Ask whether sub-agents working on your behalf are operating with the minimum permissions needed, or with a copy of everything the top-level agent has.

← Back Next → MIT 02: Isolate Agent Identities