Skip to content
be-human.ai
← Back to the journal

Human oversight

Superintelligence and Cybersecurity: Basics for Small Companies

Access, credential and logging basics a small company should fix before any more capable AI agent gets wider permissions, with a checklist and approval tiers.

Be Human typographic cover with the title: Superintelligence and Cybersecurity: Basics for Small Companies
Conceptual illustration created with AI · Be Human

For a small company, superintelligence cybersecurity starts with the same questions as any software access: which identity an agent uses, what each credential can reach, and whether every action leaves a record a person can read. Fix those three things before a more capable agent receives wider permissions. Model settings matter less than the boundaries around the agent, because those boundaries decide what a wrong instruction can actually do.

Key takeaways

  • Give each agent its own identity, never a staff member's login or a shared admin account.
  • Scope each credential to one system and one class of action, and deny anything not listed.
  • Log every tool call and approval in a store the agent cannot edit, with personal data redacted.
  • Keep sending externally, moving money, deleting data and changing permissions behind a named person.
  • Test revocation before you need it, so you know the agent stops within minutes, not days.

Why access matters more than the model

Most owners ask what an AI agent might say. The more useful question is what it can reach. An agent that reads a shared inbox, updates customer records and sends replies carries the exposure of an employee holding those logins, with one difference: it can repeat an action many times before anyone notices.

Language models can also be steered by the text they read. Instructions hidden in an email, a web page or a document can try to make an agent do something its owner never intended. The OWASP guidance on prompt injection prevention is clear that these attacks cannot be fully solved inside the prompt alone, so the defense has to sit in the surrounding system. In practice, that means limiting what the agent is allowed to do, so that a successful injection has little to work with.

For a small company, this is good news. You do not need a research team to reduce exposure. You need an accurate list of agents, clean identities, narrow credentials and a log you trust.

Give every agent its own identity

Do not let an agent operate under a founder's account, a shared mailbox login or an administrator password. When the agent uses a human login, its actions cannot be told apart from a person's, and revoking the agent means changing a password that other people also depend on. A separate identity for each agent solves both problems.

A practical baseline:

  • Name each agent identity after its job, such as the invoice reconciliation agent, not after a person.
  • Record which business owner approves the identity and what it is for.
  • Keep agent identities out of personal email accounts and personal devices.
  • List every agent identity in your asset list or risk register, so none appears without an owner.
  • Write down the tools, records and actions each identity can touch, as described in entitlement mapping for AI agents.

Scope credentials to one job

A credential should answer one question: what is this agent allowed to do in this system? OAuth 2.0 was designed around scopes for this purpose, letting a token carry only the permissions a specific task needs. Use that principle everywhere an agent connects, even where a provider offers broad access by default.

The OWASP authorization guidance adds the other half of the rule: check permissions on each request, and deny by default. A system that refuses unlisted actions and records the refusal is far easier to audit than one that allows everything and hopes for the best.

What to do this week:

  • Grant read-only scope first. Add write scope only for a named action with a named reason.
  • Use separate credentials for each system the agent touches, so one leak does not expose all of them.
  • Use short-lived tokens where the provider supports them.
  • Keep secrets out of prompts, memory files, shared notes and chat transcripts.
  • Assign a rotation owner and put the rotation on a calendar with a person's name next to it.
A narrow credential turns a bad instruction into a failed request, which is far easier to recover from than a completed one.

Log what the agent does, somewhere it cannot edit

Agents without logs are difficult to supervise and almost impossible to explain afterward. OpenTelemetry describes observability in terms of traces, which follow a single request through each step it takes, and attributes, which add context to each step. You do not need that tooling on day one, but the principle is the right one: every action should be traceable to an identity, a tool, an input summary, a result and any approval that preceded it.

A minimum useful log records the agent identity, the tool called, a short summary of the input, the outcome, the person who approved the action if approval was needed, and the time. Store these records where the agent has no write permission. If the agent can edit its own history, the log is a diary, not evidence.

Logs also contain personal data, so decide what goes in before you start collecting. Redact names, contact details and identifiers from the audit trail while keeping enough detail to reconstruct what happened. The approach is covered in PII redaction for AI agent logs.

Decide which actions need a person

Not every action needs the same level of oversight. A small set of tiers keeps the rules simple enough for staff to follow and specific enough to audit. The aim is that reading and drafting can move quickly, while actions that cannot be reversed always pass through a person.

A decision guide:

  • Reading a record the agent already has access to: the agent may proceed, and the read is logged.
  • Drafting a message, quote or document for review: the agent may proceed, and a person sends or publishes it.
  • Changing a record that can be reverted: the agent may proceed within set limits, with regular sampled review.
  • Sending something outside the company, moving money, deleting data or changing permissions: a named person approves each instance.

For the highest-risk actions, a second approver may be appropriate. The criteria are discussed in two-person approval for AI agents, and the general question of where an agent should stop is covered in where your AI agent should stop and ask.

A plan to follow before wider permissions

When a team wants an agent to take on more work, the access review should come first. The following sequence is simple enough to finish in a short working session for a single agent, then repeat for each new one:

  • Inventory the agent: its purpose, owner, identity, systems and current scopes.
  • Remove any shared or personal login the agent uses, and replace it with a dedicated identity.
  • Reduce each scope to the single system and action class the job requires.
  • Turn on logging to a store the agent cannot write to, then run one sample action and confirm it appears.
  • Write the approval tiers in plain language and tell the people who will approve actions what they are responsible for.
  • Test revocation: disable the identity, confirm the agent stops, then restore it only through the approval path. The procedure is detailed in the AI agent kill switch guide.
  • Set a review date for permissions and logs, and keep the review in the same place as the inventory.

Illustrative scenario: a support agent drafts replies from a shared inbox. At first it has read access to the inbox and write access only to draft folders. After a month, the team wants it to send routine confirmations. Before that change, the owner moves the send permission to a separate identity, limits it to a confirmation template, logs each send with the approving person, and tests disabling the identity. Nothing about the model changed, but the blast radius of a mistake became small and visible.

How Be Human handles this in practice

Be Human designs, implements and operates personalized AI infrastructure for companies, including agents that keep their identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, so access is scoped from the start rather than widened later. Managed operations add recurring permission review, monitoring, integration maintenance and incident response, which turns the checklist above into a routine instead of a one-time project. If you are not sure which task should go to an agent first, the free Be Human AI diagnostic can help you find that starting point.

Frequently asked questions

Is superintelligence cybersecurity a separate discipline from normal IT security? It is better understood as an extension of ordinary security. The same practices of identity, least privilege, logging and approval apply, but they must be applied more carefully because an agent can act faster and more often than a person. A company that already manages staff access well has most of the foundation in place.

Do we need a security specialist before letting an agent act? A specialist is useful for the highest-risk systems, but a small company can start with an inventory, dedicated identities, narrow scopes and a tested log. Those steps are concrete enough for an operations lead to complete with the vendor's help. Ask the vendor to show exactly which scopes and logs it configures for your agent.

What is the first thing to check if an agent already has access? Check whether it uses a shared or personal login, because that single fact determines how hard it is to supervise or revoke. Then list every system it can reach and the actions each permits. Remove anything the agent does not need for its current job, and make sure its actions are being recorded somewhere it cannot change.

Does logging create a privacy risk of its own? It can, because logs often capture customer messages and personal identifiers. Limit what enters the log, redact personal data where the detail is not needed, and restrict who can read the full records. Keep enough information to reconstruct each action, since a log that cannot explain what happened is of little use to anyone.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →