Insights

AI Governance

Agent security and AI governance answer different questions, so split them on one page

A governance review can approve an AI agent that is still easy to manipulate, and a hardened agent can still be doing something nobody approved. Putting one agent's governance decisions and security controls side by side shows who owns each one.

November 19, 2026 · 5 min read

Agent security and AI governance answer different questions, so split them on one page

An AI governance approval does not tell you whether an agent can be talked into doing something it should not. Neeraj Sabharwal, co-founder at Trust3 AI, makes that argument in a Forbes Technology Council piece that he reposted as an X article on November 14. His title says it directly: "Agent Security Is Not AI Governance." His point is that treating the two as the same discipline, or assuming one covers the other, is itself a risk.

I agree with him, and I think the useful next step is practical. Take one agent, write down every question someone will ask about it, and sort each question into the column where it belongs. The empty cells are the findings.

Where the line falls, in his words

Sabharwal frames the two disciplines by the question each one answers. Governance asks, in his words, "How should AI be developed, deployed and monitored to ensure responsible, ethical and compliant outcomes?" It covers policy and accountability, bias and fairness, transparency, regulatory compliance, data stewardship, and model lifecycle management, and it answers to regulators, boards, and ethics committees.

Agent security asks a different question: "Given that an AI agent is executing autonomous actions across enterprise systems, how do we ensure it cannot be exploited, manipulated or compromised?" He lists four areas under it: defense against prompt injection, least privilege at the tool and API level, verification and rollback of actions, and runtime monitoring for behavior outside expected parameters.

His example is an invoice-processing agent. Governance asks whether the model is free of bias, whether there is an audit trail, and whether the vendor is compliant. It does not ask what happens if an instruction hidden inside an invoice tells the agent to send payment somewhere else. As he puts it, "Neither answers the other."

He also assigns owners: governance to the chief AI officer or risk function, agent security to the CISO and security engineering. Smaller companies may not have both seats. The split in responsibility still holds even when one person wears both hats.

One agent, two columns

Here is the exercise applied to a single agent a HealthTech company might ship: an assistant that reads incoming patient portal messages, drafts a reply, and routes the message to the right care team queue. The governance column holds decisions. The security column holds controls that make those decisions true in practice.

| Question about the agent | Governance decision | Security control | | --- | --- | --- | | Should it exist at all? | Approved use case, named clinical sponsor, stated purpose | Not a security question | | What data may it read? | Data classes allowed (messages, demographics, no full chart), and why | Scoped credentials that enforce exactly that list | | What may it do? | Draft only; a human sends every reply | No send permission granted to the agent's identity | | What if a message tries to instruct it? | Policy that patient text is data, never instructions | Input handling and testing for injected instructions | | Who reviews mistakes? | Error review cadence, owner, and what counts as harm | Logging of every draft, route, and tool call, kept and searchable | | Is it fair across patients? | Review of routing outcomes by language and patient group | Not a security question | | What if it misbehaves at 2 a.m.? | Who has authority to turn it off | A tested way to stop it, and alerts when it acts outside its normal pattern | | What changes need re-approval? | New data, new actions, new model version | Change control that blocks deploys without that approval |

Two rows have an empty security cell, and that is fine. Fairness and the decision to build the feature at all are governance questions. Every other row needs both columns filled, and that pairing is the point. A governance decision with no control behind it is a wish. A control with no decision behind it is a guess about what the business wanted.

Why conflating them goes wrong in both directions

When a team treats governance as security, it ends up with an approved agent whose permissions were never checked against the approval. The committee said "draft only," and nobody confirmed the agent's identity lacks send rights.

When a team treats security as governance, it ends up with a well-locked agent doing something nobody signed off on. The tools are scoped, the logs are clean, and no clinical sponsor ever agreed that the agent should be routing messages in the first place.

Sabharwal makes a related point about sequencing. He argues that putting governance first and security later is the wrong order, because agents already in production are exposed now and retrofitting security is harder than building it in. He also notes that the two kinds of incidents have different remediation paths: biased outputs lead to a model audit, while a compromised agent that exfiltrates data leads to incident response and regulatory notification.

What I would do with this table

For a founder, the table is a sales document as much as an internal one. A health system security reviewer and a health system AI committee will ask different questions, often in different meetings. Showing both columns for each agent tells both groups you know the difference.

For a health system, it is an intake tool. Ask the vendor to fill in both columns for the agent they are proposing. A vendor who fills only the governance column has not done the security work. A vendor who fills only the security column has not asked what the agent should be allowed to decide.

Start with one agent. Fill in the table with the people who own each column in the room together. Then look at the cells nobody could fill. Those are this quarter's work.