PART 4
Prevention
Slides 18–28 · Nine mitigations, a coverage matrix, and a quiz
Slide 18 · Prevention Overview
Nine mitigations — each one closes a specific gap in agent identity and privilege
These aren't generic security advice. Each one targets a specific failure mode we covered in Parts 2 and 3.
📄 OWASP Top 10 for Agentic Applications 2026 · ASI03
🎯
MIT 01 — Task-Scoped, Time-Bound Permissions
Issue short-lived, narrowly scoped tokens per task. Cap rights with permission boundaries. Limits blast radius and blocks delegated abuse.
🏝️
MIT 02 — Isolate Agent Identities and Contexts
Run per-session sandboxes with separated permissions and memory. Wipe state between tasks to prevent memory-based escalation.
🔐
MIT 03 — Mandate Per-Action Authorization
Re-verify each privileged step with a centralized policy engine that checks current authority — not just session-start authority.
👩💼
MIT 04 — Human-in-the-Loop for Privilege Escalation
Require human approval for high-privilege or irreversible actions. The safety net that catches what automated controls miss.
📜
MIT 05 — Define Intent: Bind Tokens to Signed Purpose
Bind OAuth tokens to a signed intent including subject, audience, purpose, and session. Reject use where bound intent doesn't match current request.
🏢
MIT 06 — Agentic Identity Management Platforms
Treat agents as managed non-human identities with scoped credentials, audit trails, and lifecycle controls. Major platforms now support this.
🔒
MIT 07 — Bind Permissions to Context, Require Re-auth on Switch
Bind permissions to subject, resource, purpose, and duration. Re-authenticate on context switch. Auto-revoke on idle or anomaly.
📡
MIT 08 — Detect Delegated and Transitive Permissions
Monitor when agents gain permissions indirectly through delegation chains. Flag low-privilege agents inheriting high-privilege scopes.
🚨
MIT 09 — Detect Abnormal Cross-Agent Elevation and Device-Code Phishing
Monitor agents requesting new scopes or reusing tokens outside their original signed intent. Alert on device-code style flows initiated by agents.