Controlling what AI agents can change in IT and security
Give help-desk and security agents the routine tools, hold the destructive ones for a person, and keep the highest-risk ones out of reach. With NobleCloak, policy is set per tool and scoped by role, so a tier-1 assistant can reset a password but never delete an account, and a remote wipe waits for an approver. Every decision is recorded for the security team to review.
100 scenarios in 5 areas
These are scenarios for how teams govern AI agents in IT & Security Operations. Some depend on capabilities still in development. Products named in a scenario are systems your team runs, named to make the example concrete. The policy applies to the AI agent's calls into them.
Identity & Access Management20 examples
Super Admin grants need confirmation
The IT help desk agent can reset MFA factors in Okta on its own, but adding anyone to the Super Admin group pauses for an in-session confirmation, and only the IT Ops role can see that tool at all.
Read the full exampleTeller provisioning in core banking
A credit union's onboarding agent creates teller user IDs in core banking behind a confirmation gate today; routing the approval to a second operations officer for dual control is the next step.
Privileged groups trigger a pause
The joiner agent assigns Okta groups from the HRIS record automatically, and asks for approval only when the target group is flagged privileged (domain admins, payments, prod DB).
One named owner for GitHub org changes
`github.add_org_owner` is denied for everyone except a per-user override for the security lead, whose calls still require in-session confirmation.
IAM admin signs off on SSO app removal
Removing an app from Okta SSO can lock a whole department out, so the request goes to a named IAM admin, who signs off before the app is removed.
Stale account cleanup, one at a time
The hygiene agent finds 140 Okta accounts with no sign-in for 90 days and proposes deactivation; each deactivation pauses for confirmation and lands in the ledger with the evidence the agent used.
EHR security class changes at a hospital
Changing a clinician's Epic security class pauses for confirmation, and the ledger records which agent proposed it and who confirmed, for HIPAA access audits.
Tier-1 resets, never deletes
The Tier-1 support role is allowed `okta.reset_password` and `okta.unlock_user`, while `okta.delete_user` is denied for that role at the gateway, whether the request comes from Claude, Cursor, or the in-house help desk bot.
Conditional access is off-limits
No AI assistant at a regional bank may modify Entra ID conditional access policies; the tool is denied org-wide, so a mistyped request can't weaken sign-in rules for every branch.
Break-glass account stays invisible
The tool that retrieves break-glass admin credentials from the vault is not entitled to any role, so no agent can list it, discover it, or call it, even when asked directly.
Quarterly access review evidence
During a SOX access review, the audit ledger shows every Okta group change the IAM agent made, who asked for it, and who approved it, which the team hands to auditors instead of rebuilding from screenshots.
Prod key minting blocked
`aws.iam.create_access_key` is allowed on the sandbox AWS connection and denied on the production account connection, so long-lived prod keys can't be minted by an assistant.
Preview a contractor's reach
Before an external contractor's first day, the admin uses "Inspect as user" to confirm their Claude session sees GitHub read tools and nothing from Okta, AWS IAM, or the HRIS.
Sub-agents don't inherit admin
The access-request agent runs with directory-write tools, but when it spawns a sub-agent to summarize a ticket thread, that sub-agent gets only read tools; it does not inherit its parent's reach.
Two-hour elevated AWS access
An on-call engineer asks for AWS admin during an incident, and the grant expires on its own after two hours instead of lingering until someone remembers to revoke it.
MSP reads groups, cannot write
A managed service provider's identity agent may call in over A2A to read group membership for its monthly report, but write tools aren't on its peer allowlist, so it can't change memberships.
SCIM tokens stay human-only
Generating new SCIM provisioning tokens in Okta or Google Workspace is denied for all AI, since a leaked token lets an outsider create users at will.
Scripting override for one engineer
The IAM engineer who maintains Google Workspace directory automation gets a per-user allow on `directory.update_user`, while the rest of IT stays on ask.
Role change follows the directory
When an analyst moves from IT to Finance in Workday and Okta updates her role, her assistant loses the Okta admin tools at her next session, with no manual cleanup in NobleCloak.
Cap runaway password resets
A looping agent that tries to reset 500 passwords in ten minutes hits a per-user, per-tool quota and stops, leaving the help desk to look at what went wrong.
Endpoint & Device Management20 examples
Remote wipe needs a human yes
Any agent calling Intune or Jamf `wipe_device` stops for in-session confirmation, because a wrong serial number erases a working laptop.
BitLocker recovery keys at bank branches
Reading a branch workstation's BitLocker recovery key asks for confirmation and is logged with the requester and ticket, so key retrievals are traceable for examiners.
Laptop orders under a cap
The provisioning agent orders hardware from CDW and Dell only, with a $2,500 per-call cap, a monthly budget, and approval above $1,500 for anything outside the standard spec.
Ask only for the prod patch ring
The patching agent deploys Windows updates to the pilot ring freely, but deploying to the broad production ring requires approval.
Clinic kiosk reconfiguration pauses
Changing kiosk-mode settings on patient check-in iPads asks first, so an agent can't take a lobby device out of locked mode by accident.
Help desk locks, endpoint team wipes
The help desk role can lock a lost device; only the endpoint engineering role is entitled to wipe, so the wipe tool isn't even visible to help desk assistants.
Pilot Jamf instance only
Script and configuration-profile pushes are allowed against the Jamf test instance connection and denied on the production fleet connection.
No EDR exclusions from AI
Adding a path or hash exclusion in CrowdStrike is denied for every AI app, since that is exactly what an attacker would ask an agent to do.
Employees report, never remediate
The all-employees role is entitled only to "report lost device" in Intune, so staff assistants can start the process without seeing lock, wipe, or retire tools.
MDM change evidence for HIPAA
Every MDM profile change on clinic tablets, whether made by an agent or approved in-session, is in the audit ledger with inputs and approver, ready for the HIPAA security risk assessment.
Developers' Cursor can't touch Jamf
The engineering role isn't entitled to any endpoint management tools, so a developer asking Cursor to "fix my laptop config" gets nothing back from the Jamf API.
Asset reconciliation stays read-only
The finance asset-reconciliation agent can read device inventory from Intune and ServiceNow CMDB, and every write tool is denied for its role.
Drift agent opens tickets, doesn't fix
An autonomous compliance-drift agent's allow-list covers device-compliance reads and `jira.create_issue` only; it flags non-compliant laptops but cannot remediate them.
Patch only in the maintenance window
Fleet-wide restarts are callable only between 10 p.m. and 4 a.m. local, so a daytime prompt can't reboot a call center mid-shift.
Endpoint triage cost by team
Token and tool-call usage for the endpoint triage agent is metered per team, showing that one office's recurring printer issue accounts for a third of the agent's cost.
MSP remediation over A2A
The desktop MSP's agent is peer-allowlisted to call in and run approved Intune remediation scripts; every call it makes lands in the customer's ledger under the MSP's identity.
USB restrictions can't be relaxed
Changing removable-storage policy on bank teller workstations is denied for all AI, protecting the control that keeps customer data off thumb drives.
Check a new tech before first shift
The IT manager runs "Inspect as user" on a new desktop technician to confirm lock and reassign are on ask and wipe is absent before the tech's first shift.
Bulk retire capped per day
Retiring more than 25 devices in a day hits a quota, so a bad CSV can't decommission a whole floor's laptops at once.
Software installs from the approved catalog
The self-service agent can install from the company portal catalog, while paid license purchases go through spending policy limited to the approved software vendors.
Security Operations & Threat Response20 examples
Host containment for analysts only
`crowdstrike.contain_host` is on ask for the SOC analyst role and not visible to Tier-1, so containment always has a qualified person confirming it.
Suspend a compromised user
When the SOC agent concludes an account is compromised, `okta.suspend_user` pauses for confirmation, and the ledger links the suspension to the alert that triggered it.
Firewall blocks go to the network lead
The agent proposes a Palo Alto block for a malicious IP, and the approval routes to the network lead instead of whoever happens to be chatting with the agent.
Phishing purge across mailboxes
The phishing triage agent reads reported emails freely, but purging a message from every M365 mailbox asks first, since a false positive deletes legitimate mail company-wide.
MDR partner requests, we confirm
The MDR provider's agent may call in over A2A to read alerts and request containment, but containment on our side still requires an in-session confirmation from our analyst.
Acknowledge yes, resolve asks
The SOC agent can acknowledge PagerDuty incidents on its own, while resolving one asks first so incidents aren't closed before a human has looked.
Prod vulnerability scans ask
Launching a Tenable scan on the staging connection is allowed; launching one on the production network connection asks, since aggressive scans have tipped over legacy systems before.
Credential-stuffing response at a credit union
During a credential-stuffing attack, the agent can force-reset members' online banking credentials behind confirmation, with every reset in the ledger for the NCUA incident report.
Dual control for mass suspension
Suspending more than ten accounts at once needs sign-off from two SOC leads, so one compromised session can't lock out the entire company.
Search freely, never delete indexes
The investigation agent can run any Splunk search, but deleting saved searches, indexes, or alerts is denied, keeping the evidence base intact.
Threat intel only through the proxy
Lookups to VirusTotal and the paid intel feed go through the governed egress proxy; the agent can't fetch arbitrary URLs it pulled out of a malicious email.
Sandbox detonation on a budget
The paid malware sandbox API has a per-call cap and a monthly budget, so an alert storm can't run up a five-figure bill overnight.
Who authorized containment
In the post-incident review, the ledger shows the order of events: which agent proposed isolating each host, who confirmed, and the inputs it saw.
Forensics sub-agent gets read-only
The incident agent spawns a forensics sub-agent to pull process trees; the sub-agent does not inherit containment or account-suspension tools.
Logging can't be switched off
Stopping CloudTrail, deleting log groups, or disabling Microsoft Defender audit logging is denied for every agent, covering the first thing an attacker does after a takeover.
Prompt-injected exfil attempt blocked
A phishing email tells the triage agent to post mailbox contents to a pastebin; the egress proxy refuses the call and the denied attempt shows up in the ledger for the SOC.
Secrets scrubbed before the model
API keys and passwords in raw alert payloads are redacted before they reach the LLM, so the model provider never sees live credentials.
GuardDuty findings can't be archived
The hunting agent can read and annotate GuardDuty findings but `archive_findings` is denied, so findings can't disappear from the queue before a human reviews them.
Investigation cost per incident
Token and tool-call cost is attributed to each incident's agent session, giving the CISO a real figure for what AI-assisted triage costs per severity level.
Containment only inside the playbook
The containment tool is callable only from within the approved SOAR ransomware playbook, not from a freeform chat.
SaaS Administration20 examples
Drive ownership transfer asks
Transferring a departing employee's Google Drive files to a manager pauses for confirmation, so files don't land with the wrong person.
True-up approval over threshold
An annual license true-up above $10,000 requires approval before the agent submits the order, and anything smaller goes through automatically.
Jira projects can't be deleted
Deleting a Jira project is denied, and changing Confluence space permissions asks first, protecting years of tickets and documentation from one bad prompt.
External sharing toggles ask
Turning on external link sharing for a Google Drive or Box folder asks for approval when the target domain is outside the company.
Branch protection for platform team only
Editing GitHub branch protection rules is visible only to the platform engineering role and still asks first for them.
Salesforce admin gets ask, not deny
The senior Salesforce admin has a per-user override that moves production permission changes from deny to ask, while everyone else stays blocked.
Mailbox delegation at a health system
Granting full access to another person's M365 mailbox asks first and is logged, since clinical mailboxes contain PHI.
Slack exports stay human
The IT agent can create Slack channels and archive inactive ones, while full workspace export is denied, since an export is a copy of every message the company has sent.
Salesforce permissions only in sandbox
Profile and permission set changes are allowed on the Salesforce sandbox connection and denied on production for every AI app.
License purchases from approved vendors
The SaaS admin agent can buy additional Zoom, Slack, and Atlassian seats, limited to those vendors, with a monthly budget and a per-call cap.
No surprise SaaS signups
An agent asked to "find a tool for this" can't sign up for new SaaS: the egress proxy blocks unknown sites and spending policy rejects vendors outside the allowlist.
Repos never go public
Changing a GitHub repository's visibility to public is denied for every AI, so private code can't be exposed by an agent misreading a request.
IT doesn't touch Workday config
The IT agent isn't entitled to Workday tenant configuration tools at all, because HR owns that system and its change process.
Snowflake cost by team
Usage attribution shows which teams' agents drive the most Snowflake MCP queries, so FinOps can set chargebacks on real numbers.
OAuth consent grants denied
Granting OAuth consent to third-party apps in Google Admin is denied for all AI, closing a common route for data-harvesting apps.
Create changes, never approve them
The agent can draft ServiceNow change requests with full detail, but `approve_change` is denied, keeping the change advisory board's sign-off in human hands.
SOC 2 change evidence
Every SaaS configuration change made through an agent is in the ledger with requester, agent, inputs, and approval, which becomes the population for the SOC 2 change-management sample.
Bulk license removal quota
Removing more than 50 licenses per day is quota-limited, so an agent misreading a spreadsheet can't strip a sales team of Salesforce mid-quarter.
Consultant's agent stays in sandbox
The Salesforce implementation partner's agent is peer-allowlisted for A2A calls into the sandbox connection only; production is not reachable on its delegation.
Network, Cloud & Backup20 examples
Terminate instances: staging free, prod asks
`ec2.terminate_instances` is allowed on the staging AWS connection and on ask for the production connection.
Restores ask and leave a record
Restoring a database from backup pauses for confirmation, and the ledger captures which snapshot, which target, and who approved.
Reserved instance purchases
The FinOps agent can buy AWS Savings Plans and reserved instances under a per-call cap, with approval required above $10,000.
DNS record edits ask, zones never delete
Route 53 and Cloudflare record changes ask for confirmation; deleting a hosted zone is denied outright.
Snowflake drops denied, resizes ask
`DROP TABLE` is denied on the Snowflake production connection, and warehouse resizes ask first because a 4XL left running costs real money.
Renew certificates, ask before revoking
The certificate agent renews expiring TLS certificates on its own, while revocation asks first because revoking the wrong certificate takes down a customer-facing site.
Plan yes, apply no
Platform engineers' assistants can run `terraform plan` against production, but `terraform apply` on the prod workspace is denied; applies go through the pipeline.
Block open-to-the-world rules
A security group change that opens a port to 0.0.0.0/0 is denied, while scoped CIDR changes are allowed.
No public S3 buckets
Setting a public ACL or disabling Block Public Access on any S3 bucket is denied for every AI, whichever account connection is in use.
Recovery points can't be deleted
Deleting AWS Backup or Veeam recovery points is denied for all agents, so a compromised session can't wipe the backups before a ransomware strike.
Read the full exampleCore banking segment locked
Firewall rule changes on the core banking network segment are denied for all AI, while branch office network changes are on ask for the network team.
No namespace deletes in prod
`kubectl delete namespace` is denied on the production cluster connection, while Cursor in the dev cluster can create and tear down namespaces freely.
Prod Postgres is read-only
The analytics assistant can run SELECT queries on the production Postgres connection, but write and DDL tools are denied there and allowed on staging.
Packages only through the proxy
Cloud agents can't curl arbitrary install scripts; egress is limited to the governed proxy in front of approved package registries.
FinOps agent pays its own way
The FinOps agent's token and tool-call cost is metered alongside the savings it finds, so leadership can see whether it covers its cost.
VPN access for the network team
VPN user provisioning in the remote-access gateway is visible only to the network operations role, so general IT assistants can't hand out remote access.
Change freeze enforced
During the year-end change freeze, infrastructure write tools become ask for everyone and deny for autonomous agents, then revert automatically on January 3.
DR drill agent and its sub-agents
The disaster-recovery drill agent's allow-list covers failover tools for the DR region only, and the sub-agents it spawns for health checks get read access, not failover.
Backup testing evidence for examiners
The ledger shows every quarterly restore test the backup agent ran, with snapshot IDs and outcomes, giving FFIEC examiners evidence of tested backups.
One MSP, many credit unions
An MSP serving several credit unions runs each as its own NobleCloak org; tenant isolation keeps one credit union's cloud policies, ledgers, and data separate from the others.
| Action | Who | Decision |
|---|---|---|
| reset_password | help-desk agent | allowed |
| delete_user | any agent | denied |
| grant_admin_role | IT admin's assistant | needs approval |
| remote_wipe_device | endpoint team's assistant | needs approval |
| edit_conditional_access | any agent | denied |
| search_logs | SOC analyst's assistant | allowed |
Approval requests go to the identity or endpoint lead in Slack or the web inbox (Teams and email coming soon).
Illustrative: names no package, import or API call
AI agents in IT & Security Operations: common questions
Can an AI agent grant admin rights?
Only with approval, if that is how you set it. Marking the admin-grant tool as needs approval pauses the run until the identity lead answers; denying it keeps it away from every agent.
How do the help desk and the endpoint team get different tools?
Policy is scoped by role, so the help desk's assistants can lock a device while only the endpoint team's can request a wipe. Inspect as user shows the tools a given person's agent would get.
Can an analyst contain a host without waiting for approval?
That is your call per tool and per role. Some teams allow containment for analysts and require approval for everyone else; others hold every containment for the SOC lead.
Can elevated access expire on its own after an incident?
Time-boxed grants are coming soon, and the scenarios that need them carry the Coming soon label. Until then elevated tools are granted by role and removed by changing the policy.
Set this policy up for your own agents.
Start with your own agents, your tools and the approvers you already have.
Other functions
Finance & Accounting
Payments, the close, treasury, budgets and payroll.
Software Development & Engineering
Code review, releases, production systems and on-call.
People & HR
Hiring, onboarding, compensation, employee relations and training.
Sales & Revenue
Outreach, the CRM, quotes, contracts and renewals.
Marketing & Communications
Content, campaigns, paid media, social and the website.
Customer Support & Success
Ticket replies, refunds, privacy requests and escalations.
Legal, Risk & Compliance
Contracts, regulatory work, audit evidence, vendor risk and records.
Operations, Procurement & Facilities
Purchasing, vendors, inventory, facilities and projects.
Data, Analytics & AI Operations
Warehouse access, pipelines, agent-to-agent work, models and AI cost.

