problem_kicker
An agent that can call tools is an actor in the security model. Giving it a broad API key or inheriting a human session makes authorization ambiguous, difficult to revoke and hard to audit.
DEMAND LANGUAGE / REAL-WORLD PROBLEM
“Can AI do this itself without getting full access to everything?”
“What happens when the agent makes the wrong change?”
WHAT CAUSES THIS?
Shared credentials erase actor identity.
architecture_for SECURE AI AGENT PERMISSIONS
We inventory machine actors, tools, data boundaries and irreversible actions. Each capability receives explicit policy, credential lifetime, approval requirements and post-condition verification.
Prefer short-lived credentials, least privilege, deny-by-default policies, human approval gates for high-impact actions, dependency inventories and tamper-evident audit evidence.
Authorization must be observable but should not force every action through a high-latency remote path. Cache only policy decisions whose inputs and expiry semantics are explicit.
AI agents · IAM · policy engine · short-lived credentials · audit logs
failure_kicker
measure_kicker
verify_intro
CTO / CIO FAQ
No. A tool may expose multiple resources and actions. Authorization normally needs actor, capability, resource, context and policy.
Only where the delegated-user model is intentional. Autonomous machine work is usually clearer with a distinct workload identity and scoped delegation.
At the boundary before a high-impact capability is authorized, with enough context for a human to understand the proposed change.
connected
