The essence is that any agent functions as an account in the system, accessing data through specific credentials with corresponding permissions. This fundamental question has been discussed for thirty years, but modern systems process it at machine speed.
Systems can show which agents exist and what they are capable of reading, modifying, or deleting in the database. However, the decision to grant these permissions was often made unintentionally, simply during the setup process.
The number of non-human identifiers now exceeds the number of people: on average, enterprises have 45 times more such accounts, and in cloud environments—144 times more. Furthermore, half of organizations cannot name the owner of these agents, and many use agents layered over existing service accounts that no one has audited.
Instructions given to an agent are not access control. For example, in April, a production database and its backups were deleted in nine seconds due to a token that had such authority, despite the agent's instructions forbidding such an action. The decisive factor was precisely the permission assigned to the account in the database.
Four questions for assessment
The need for a new platform does not arise if four key questions can be answered: which non-human identifiers have access to each database and who is responsible for them; what each of them can do besides read; when each access expires; and who reviews the audit logs.
At the Microsoft Ignite conference, taking place in San Francisco on November 17, the topic of agents is central. Microsoft asserts that agents are 'your newest insiders,' possessing real access to data, identities, and the cloud, and that any decision to implement them is a trust decision. Sessions cover the readiness of access control systems to work with agents at scale and the ability to detect agents running on employees' laptops.
The discussion about managing AI agents is happening at too high a level. At the database level, an agent is not a new phenomenon, but simply an account. Whether it connects directly or via an application, it reaches the core of the system as an account, with the same rights as a person at the keyboard. The novelty is the speed, not the principle itself, and most companies have never answered this question for existing accounts.
The situation in South Africa is not ideal. The information regulator received 1,220 security breach notifications in the first five months of the fiscal year, leading to a forecast of over 3,000. These incidents are linked to weak passwords, human error, and 'insufficient security controls.' A Sophos study showed that in 85% of ransomware attacks, the most serious organizational identity breach coincided.
Credentials are already being used to exfiltrate data from the country, and an agent is just another such account. In October, a pathology lab reported an attack where the attacker claimed to have copied about 375 databases, including one admission table with over 80 million records. Although the group noted that these were records, not patients, and the application vulnerability was patched, this volume of data is typical for an organization of this size.
Newest Insiders
Microsoft has already described what will be presented at Ignite. In October, the company's identity guide stated that an agent should start without any privileges and only receive what is necessary for the approved task, with clear time and scope boundaries, leaving high-impact actions to a verified human. An agent does not inherit the permissions of the employee sponsoring it. It must be assumed that it could fail or be compromised, and controls must be built independent of the model's behavior.
A practical guide published by Microsoft in July is more detailed: every agent should be treated as an independent entity with a dedicated identity, not as a shared secret or a reusable service account. It is necessary to separate what an agent can read from what it can write. Temporary privilege elevation is preferable to permanent access, as 'temporary access without an expiration mechanism practically becomes permanent.'
It is important to log the agent, its scope, and the human for whom it acted, because 'logging only the LLM response creates an audit trail that looks present but is useless for forensics.'
All these requirements relate to permission management: account, role, grant, expiry, audit. None of this is new to those who managed databases. However, it has not been implemented in most systems.
One preliminary review of Ignite pointed out the problem accurately: 'The agent registry is the easy part; deciding what each agent can see is the hard work.' This decision is made where the data resides, by the person holding the keys to the system core, and often this was done years ago for completely different purposes.
Cloud Security Alliance collected statistics in May: non-human identifiers—service accounts, API keys, integration accounts, and now agents—exceed the number of people in an average enterprise by 45 times, and in cloud environments—by 144 times. One out of twenty has full administrative rights. Half of organizations cannot say who owns their AI or agents, and 78% lack a documented process for creating or decommissioning them. The study summary states: 'The agent disappears; the credentials remain.'
Agents are quickly joining this population: according to Microsoft itself, active agents in Microsoft 365 have increased 15 times in a year. Gartner expects that by 2027, 40% of enterprises will either downgrade or decommission autonomous agents due to governance gaps discovered only after a workplace incident. This is how permission issues are most often found.
Nine Seconds
The benchmark incident of the year is small, well-documented, and entirely related to access rights. In April, a startup's coding agent encountered a credential mismatch in a test environment and began fixing the issue. It discovered an API token in an unrelated file, 'assumed'—as it later said—that deleting the storage would be limited to the test environment, and triggered the deletion endpoint. This storage contained a production database and its backups. The founder noted: 'It took nine seconds.'
The token was created for managing user domains, and the platform offered no way to narrow its authority, so it could do anything. Despite the agent's instructions—the system prompt tool and project rules—forbidding any destructive actions without a request, it proceeded, acknowledging this. The CEO of the hosting provider, with the honesty he respects, stated: 'If you (or your agent) authenticate and call the delete command, we will execute this request.' The data was restored within an hour after management intervention, and the outdated endpoint called by the agent, which lacked protection against delayed deletion, was fixed within days.
This lesson has already been learned. The prompt is the policy the model must enforce. Permission is the policy the system enforces for you constantly, regardless of speed. When they diverge, permission wins, so management must reside there. This place is not the only one. Critical permission resides where the data can be accessed: in the database account, in the identifier the application connects with, or in the platform token that can delete the core data volume. The database is the last of these points, and it is what most companies have never checked.
The author has seen similar situations for years, even without agents involved. An application account created in 2014 as db_owner by installer requirement still holds such rights in a database containing all customer numbers. A scheduled task running as sysadmin because it always worked belongs to a consultant who left in 2019. A service account whose password is stored in a configuration file accessible to the entire company is used by three systems, one of which has been decommissioned.
By pointing an agent at such a system, you are not introducing a new risk; you are simply giving a round-the-clock user privileges that no one has checked.
There is a correct model, and it is not exotic. The best example the author saw this year was how a database administrator created an agent to diagnose a SQL Server system: an account with view-server-state and read-only rights, a list of permitted operators that must pass before entering the core, and one constant instruction: 'You discover and report; you never fix.' In this design, all controls are executed outside the model: the account, the list of permitted operations, the human standing in the way of any state changes. This is the lesson of April, built from the ground up.
On Whose Behalf
Permission also defines the second aspect—accountability. Here, the agent complicates the situation. When an agent acts under delegated authority from a human, that human appears in the log. If it acts under a general service account, the account itself appears in the log. In both cases, nobody subsequently knows whether the human or the software decided it, although Microsoft leadership insists that the audit trail must maintain this distinction.
King V asks the board of directors about human oversight and the ability to reverse AI decisions. Popia is more direct: the organization, not the vendor, the model, or the agent, is responsible for processing.
After an incident, people ask not what the software recommended, but who approved the action it took. And in the database, this answer is a grant with a specified name, or there is none.
Gartner's four levels of agent autonomy—observation, consultation, action with human approval, autonomous action—almost perfectly map to database privileges: read-only; read and suggest; write with human involvement; free write. Most systems the author analyzed provide the fourth level for what required the first. The quickest test is to list every non-human account and ask what it can delete.
Four Questions
None of this awaits a platform announcement. A board of directors or committee to which it delegates authority can ask four questions this quarter and expect answers within a month.
Which non-human identifiers can access each database containing personal information, and who is responsible for them? This list will include service accounts, integration accounts, scheduled jobs, and reporting connections before any agent appears, and the owner column will be blank.
What can each of them change, not just read? Any non-human entity with sysadmin, db_owner, or equivalent rights requires written justification, and most will not have it.
When does each access expire? Permanent access without an end date is permanent access in practice, according to Microsoft. Agents intended to have task-limited credentials can be better managed than integrations set up ten years ago.
Who reviews the log, and does it record on whose behalf the action was taken? A log showing only the account does not answer the question that an incident will ask.
Ignite will present registries, conditional access for agents, behavior-dependent risk assessments, and sandboxes that keep controls out of the model's reach. Each of these tools assumes that someone correctly set the permission in the database at the end of the chain. This someone is the client, and the work does not wait for the keynote speech.
Agents are already in the data. The question in eighteen months is what they were allowed to do for each of them, and whether that matched what they could actually do.
Concluding Remarks
The author spent most of his career working with database systems, and the pattern hasn't changed in 30 years: access was granted hastily and never checked. Agents will not change this pattern; they will execute it at machine speed. Organizations that treat every agent as a database user—which it is: an account with an owner, scope, expiry, and log—will be able to adopt...
