CyberPrism.App

Identify each agent's owner, authentication method, connected tools, readable data, and permitted actions. Prioritize agents that can change production systems, export sensitive data, or grant access; require evidence that their activity is traceable and their access can be revoked.

As of October 12, 2026, Microsoft's October 1 Digital Defense Report offers a timely reason for this review: agent security depends on surrounding identities, permissions, tools, and infrastructure. It highlights attribution and access revocation alongside agent identity. [Microsoft's report overview](https://www.microsoft.com/en-us/security/blog/2026/10/01/insights-from-the-2026-microsoft-digital-defense-report/).

The checklist below is a suggested operational review, not a Microsoft compliance standard.

Record the agent name, business purpose, accountable owner, runtime, identity, credential location, connected services, permission scopes, and last review date. Distinguish its own credentials from access delegated by a person or supplied by a shared integration.

Trace one representative task through every tool it calls. If a connector uses a separate account, include that account's permissions: the agent's visible settings may describe only part of its effective access.

For each connection, separate reading, writing, deleting, external sharing, and permission administration. Compare those actions with the approved business task and document why each sensitive permission is necessary.

For example, an agent preparing incident summaries may need selected incident records but no ability to close tickets or change responders' access. Test narrower permissions with representative tasks before applying them in production.

Suggested escalation rule: suspend unowned privileged connections pending review. For owned connections with excessive access, assign a reduction deadline and require explicit approval for any temporary exception.

Run a harmless test against a designated test resource. Check whether retained logs connect the initiating request, agent identity, tool action, target resource, timestamp, and outcome; record any gaps that would obstruct an investigation.

In an approved test environment, suspend the agent and revoke the relevant access using the connected service's controls. Check existing sessions, queued jobs, retries, and alternate credentials rather than assuming a stopped interface ends all activity.

Record the observed result and any delay before access stops. Credential rotation alone is insufficient evidence that previously issued sessions or tokens have lost access.

For each agent, choose retain, restrict, or suspend. Attach the owner, sensitive resources reachable, permission changes, logging gaps, revocation test result, and next review date.

Keep the decision specific: 'Restrict the support agent to approved queues; remove external sharing; owner to verify by October 19.' This gives leadership an action and a completion condition.

Repeat the review when the agent gains a tool, changes identity, receives broader permissions, or starts a new business task. A previous approval covers the access reviewed at that time.