Personalized AI infrastructure
Superintelligence Risk Register: Ten Starter Entries With Owners
Ten starter entries for a risk register covering AI agents, each with a named owner and a review trigger, that a business can adapt this week.

A superintelligence risk register is a short, living list of the ways a capable AI agent could harm your company, with one named owner and one clear review trigger for each risk. For most companies, the practical starting point is ten entries covering scope, memory, permissions, external actions, inputs, data, quality, cost, dependence and incident response. The term superintelligence is used loosely in business, so each entry below describes something you can observe in an agent you run or plan to run today, and it can be scaled up as capability grows.
Key takeaways
- Write each risk as an event you could observe in your own agents, not as an abstract future scenario.
- Give every entry one accountable business owner and one review trigger that fires on a visible event.
- Start with ten entries, then merge or drop them after the first review instead of expanding early.
- Attach each control to a practice you already run: permission reviews, change requests, backups and on-call paging.
- Review the register when an agent's tools, data scope, model or people change, not only on the calendar.
What a risk register is for, and what it is not
A risk register is a decision tool. It tells a leader which risks exist, who answers for each one, what control is in place and when the risk must be looked at again. It is not a policy document, a forecast of how advanced AI will develop, or a substitute for a security or legal review. The NIST AI Risk Management Framework is a useful reference for the overall structure: it groups work into governing, mapping, measuring and managing risk, and a register is the working record that connects those four activities to specific agents.
A useful register has six columns: the risk, what it looks like in practice, the owner, the existing control, the review trigger, and the evidence the owner will look at. If a column cannot be filled in, the entry is not yet actionable and should be rewritten before it is accepted.
How to write an entry someone can act on
Most risk registers fail because entries are too vague to act on. Use this checklist for each entry before you add it:
- Name the agent role the entry applies to, so the risk is attached to something specific.
- Describe one observable failure in one sentence, such as an action taken without approval or a fact repeated after it was corrected.
- Name one owner by role, and confirm that person accepts the duty in writing or in the meeting notes.
- Record the control that already exists, even if it is manual, such as a human approval step or a monthly log check.
- Set a review trigger that someone can detect, and state where the evidence will be found.
- Mark the date of the next scheduled review, so the entry cannot sit unexamined.
The ten starter entries
Each line below gives the risk, its owner and its review trigger. Adapt the wording to your company, and keep the first version to one page.
- Scope creep: the agent starts taking on tasks outside its defined role. Owner: the business sponsor of that role. Review trigger: any new task type, tool or channel added without an approved change request.
- Memory drift: stored facts become outdated or wrong, and the agent keeps acting on them. Owner: the operations lead responsible for the source records. Review trigger: a corrected fact that the agent repeated, or memory older than the refresh window the team agreed on.
- Over-broad permissions: tools are granted beyond what the role needs, often because access was added for a one-off task. Owner: the IT or systems lead. Review trigger: the scheduled permission review date, or a change in the role or employment status of the person who approved the access.
- Unapproved external action: the agent sends a message, changes a record or triggers a payment without the required approval. Owner: the process owner of the affected workflow. Review trigger: any logged action that falls outside the approval boundary.
- Instructions hidden in inputs: text inside emails, documents or web pages tries to redirect the agent. Owner: the security lead. Review trigger: any flagged instruction in tool inputs, or the addition of a new external input source.
- Confidential data leaving its boundary: know-how or personal data appears in prompts, memory or third-party tools where it should not. Owner: the person responsible for data protection in the company. Review trigger: a new vendor, a new category of data or a change in where data is stored.
- Silent quality decline: outputs get worse after a model, prompt or tool change, and nobody notices because no one measures them. Owner: the business owner of the role. Review trigger: a failed scheduled evaluation, or any change to the underlying model.
- Cost growth without value: usage rises faster than the work the agent completes. Owner: the budget owner named in the business case. Review trigger: usage crossing the agreed threshold, or cost per completed task rising across two consecutive reviews.
- Single-point dependence: only one vendor or one person can operate, repair or restore the agent. Owner: the operations leader. Review trigger: a contract renewal, a staff change, or a failed restore test.
- Unowned incidents: an error happens and no one is alerted or empowered to respond in time. Owner: the named on-call lead. Review trigger: any incident that passes the agreed response window, or any incident with no named responder.
Assigning owners without building a committee
The most common objection to a risk register is that it needs a new committee. It does not. Each entry needs one owner, and that owner is usually someone already responsible for the workflow the agent touches. A single person can own several entries, provided the register shows that clearly. The owner should be a business or operations person, not only a technical contact, because the decisions in the register are about whether an outcome is acceptable.
If ownership is unclear for an entry, treat that as a finding. Either the agent is operating in a place nobody is accountable for, or the role has been defined too broadly. Both situations are easier to correct before the agent is widened. The steering question is simple: who can stop this agent, and who can approve its next change?
Review triggers: events first, calendar second
A quarterly calendar review is a reasonable default, but it is too slow for the events that matter most. A review trigger is a specific, observable event that moves the entry to the front of the agenda. Good triggers include a new tool, a new data source, a model change, a permission granted outside the normal process, a failed evaluation, a restore test that does not pass, or a new person joining a role the agent supports. Changes of this kind should flow through a documented request process, which is described in detail in How to Request and Approve Changes to a Managed AI Agent.
Illustrative scenario: a customer support team has an agent that drafts replies and looks up order status. The team adds read access to the billing system so the agent can answer refund questions faster. Nobody updates the register. Under the entries above, this event triggers three reviews at once: scope creep, over-broad permissions and confidential data leaving its boundary. The owners compare the new access with the role definition, remove the billing write capability that was granted by default, and record the decision. The review took under an hour because the trigger was defined before the change happened.
How Be Human's managed approach handles the register
Be Human designs and operates AI agents with the register built into the operating layer. Each agent's role, memory, tools, permissions and approval boundaries are written down before it runs, then monitored and reviewed on a schedule and on defined triggers. Managed operations include monitoring, recurring evaluations, permission review, incident response and controlled changes, so the owner named in each entry works from logs and evaluation results rather than from memory. A first engagement usually starts with one useful role, which keeps the register short enough to maintain and makes each owner's responsibility concrete from the start.
What to do this week
Start with a single agent or the one you are planning to deploy next. Copy the ten entries into a shared document, remove the ones that do not apply, and rewrite the rest using the checklist above. Name an owner for every line, including the lines you think are minor. Schedule the first review within the month, and add the review triggers to the calendar or to the change process where your team already works. Bring the document to the first review meeting with two questions: which entries lacked a real owner, and which control turned out to be missing. Those answers will shape the second version more than any further reading. Before deciding where to begin, take the free Be Human AI diagnostic to identify the first task worth handing to an agent and the risks that task would carry.
Frequently asked questions
How many entries should a starter superintelligence risk register have? Ten is a practical size for one agent role because it covers scope, memory, permissions, external actions, inputs, data, quality, cost, dependence and response. Expand only after the first review shows which entries are noise and which risks are missing. Merge entries that share an owner and a control so the register stays readable.
Who should own the risk register itself? The register should be owned by an operations leader or the executive sponsor of the AI project, because they can make decisions across functions. Each individual entry still needs its own named owner, usually the person responsible for the workflow the agent touches. Ownership of the document and ownership of the risks should be stated separately.
How often should the register be reviewed? Set a calendar review, quarterly is a common choice, and treat event triggers as the faster mechanism. A new tool, a model change, a permission granted outside the normal process or an incident should reopen the affected entries immediately. Record the date of each review and the decision reached, so the history is visible to anyone who inherits the agent.
Does a risk register replace a security or legal review? No. The register records which risks exist, who owns them and what control applies, and it tells you when a deeper review is needed. Security, privacy and legal reviews remain separate activities, and some entries, such as confidential data leaving its boundary, should trigger them directly. The register makes those reviews faster because the scope and the owner are already defined.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


