AI agents change the security conversation because they do something ordinary generative AI generally does not: they act.
A chatbot can give someone a bad answer and leave the human to decide what happens next. An agent can receive a goal, choose tools, call APIs, write data, trigger workflows, maintain state and continue working through multiple steps before a person sees the result.
That difference sounds technical until the agent has credentials.
At that point, the CISO is no longer evaluating only whether an AI system produces reliable information. The organization is deciding what authority software should be allowed to exercise, which humans remain accountable for that authority, where intervention should occur, and what happens when the agent encounters conditions nobody anticipated during design.
Microsoft's AI Agent Shared Responsibility Model is a useful starting point because it treats autonomy as a change in the operating model rather than merely another feature of an LLM. Agents create additional responsibilities around orchestration, tools, permissions, memory, identity and autonomous action, while deploying organizations continue to retain responsibility for data, access, authorization, human oversight and accountability.
For CISOs, that means the most useful pre-deployment questions are not just about the model.
They are about the system of authority around the agent.
1. What is this AI agent actually allowed to do?
This should be one of the first questions because “agent” can describe very different things.
One agent may search internal documents and draft a recommendation. Another may update customer records, send external communications, execute code or change production infrastructure. Those systems should not inherit the same security assumptions simply because they use similar underlying AI.
The useful inventory is therefore not merely a list of agents. CISOs need to understand the actions available to each agent, the tools it can call, the systems those tools reach and the consequences those actions can create.
Microsoft's current guidance makes this distinction explicit by separating the tools and actions layer from the underlying model. It recommends least privilege for each tool, authorization at the action level and stronger human controls around sensitive or difficult-to-reverse activities such as payments, production changes, deletes and external communications.
The risk of an agent is shaped at least as much by what it can do as by what it can say.
2. Whose authority is the agent using?
This question becomes complicated quickly because an agent can act under several different authority models.
It may operate using its own identity, use delegated permissions from the person who invoked it, or combine those patterns depending on the task. Microsoft now distinguishes between agents acting as themselves and agents acting on behalf of users because the identity relationship determines which permissions and access rules apply.
For a CISO, the practical concern is whether that relationship is understandable and enforceable.
If the agent uses a broad standing identity, could it perform actions the requesting employee would never have been allowed to perform directly? If it acts on behalf of a person, are permissions constrained to what that person can legitimately delegate? Can the security team later determine which human, agent and credential were involved in a consequential action?
The familiar principle of least privilege still applies. Agents simply create another actor through which privilege can be exercised.
This is also where the Human-Agent Accountability model becomes useful, because authority should remain traceable as it moves from the organization, through a human role, into an agent and potentially onward into other systems.
3. How much autonomy are we actually granting?
Organizations can spend considerable time debating whether a system is technically “agentic” while avoiding the more useful question of how much independent authority it has received.
Can the agent choose which tools to use? Can it decide the order of operations? Can it initiate actions without approval, delegate tasks to another agent or continue working after the original user has moved on to something else?
Each additional degree of autonomy changes the supervisory problem.
Microsoft's shared-responsibility guidance describes agents as systems that can plan and loop through intermediate steps, use tools and APIs, hold state and interact with other agents. Controls such as step limits, iteration limits, cost ceilings and restrictions on tool chaining may look like technical implementation details, but they reflect a much broader risk decision.
Before deployment, the organization should be able to explain what the agent may decide independently and which decisions remain outside its mandate.
If that boundary cannot be described clearly, it will be rather difficult to supervise later.
4. What happens when the agent encounters something outside its mandate?
Happy-path workflows are generally impressive in demonstrations. The more revealing test is what happens when the situation stops looking like the demonstration.
The agent encounters contradictory instructions, needs access it was not originally given, discovers a customer case that does not fit the expected pattern or finds another route to accomplishing its objective. The important question is not whether the agent can improvise, but whether it should.
NIST's October 2026 work on agent identity management proposes mandates and constraints as part of the answer. The originator of an agent task should be able to define intended outcomes and boundaries, while risk thresholds can trigger human review when the system reaches conditions that should not be resolved autonomously.
Cybermaniacs approaches the same problem through Human-Agent Accountability, Intervention Points and intervention triggers.
The agent needs a mandate, but the organization also needs to know what happens when that mandate runs out.
5. Which actions require human approval or intervention?
The useful answer is unlikely to be “all of them.”
If every tool call requires a person, the agent will spend much of its life waiting for permission and the organization will have recreated a manual workflow with more interesting infrastructure. If no actions require human intervention, the organization is assuming that technical constraints and autonomous decision-making are sufficient across every condition the agent may encounter.
The sensible middle depends on consequence.
Microsoft recommends human-in-the-loop gates for high-impact, irreversible and sensitive actions. NIST similarly recommends organizational risk thresholds that trigger human review and warns about consent fatigue when escalation occurs so frequently that people begin approving requests habitually.
This is why Cybermaniacs treats intervention as a design problem rather than a checkbox. Our Guide to What Should Trigger Human Intervention in AI Systems? looks at conditions such as rising consequence, changing authority, uncertainty, irreversibility and mandate drift.
The CISO should be able to ask a very practical question before deployment: what can this agent do without a person, and what specifically causes that autonomy to end?
6. Who is accountable when the agent acts?
Naming an owner is useful, but it is not the same thing as building accountability.
An agent owner may approve deployment while a process manager supervises the work, an infrastructure team manages identity, a business user provides the goal and a different person receives an escalation when something goes wrong. That can be a perfectly reasonable operating model, but somebody needs to understand how those responsibilities connect.
Microsoft's shared-responsibility model states that organizations continue to retain responsibility for human oversight and for determining who is accountable for autonomous agent behavior.
At Cybermaniacs, we use Human-Agent Accountability for the wider system of human ownership, delegated authority, supervision, intervention and evidence that keeps agent activity connected to meaningful human responsibility.
The pre-deployment question is therefore not simply “Who owns the agent?” It is whether that person—or the surrounding accountability chain—has enough visibility, authority and operating control to own the outcome in practice.
Our Guide to How Do You Assign Accountability for AI Agent Actions? takes that implementation problem further.
7. Can the accountable human actually supervise the agent?
This is where workforce design starts colliding with security architecture.
A manager may formally remain accountable for an agent while having very little visibility into what happens between receiving the task and presenting the result. Another supervisor may have excellent observability but be responsible for enough agents that meaningful review becomes unrealistic.
Neither condition is fixed by adding the person's name to a governance register.
Cybermaniacs uses Supervision Sufficiency to describe whether the human supervisory capacity surrounding an AI workflow is adequate for the autonomy, complexity and consequence involved.
Before deployment, CISOs should understand what the supervisor will actually see, which conditions require attention, what expertise the role depends on and how much activity the person can reasonably handle.
This becomes especially important as agent throughput scales. Human supervision can remain technically present while the amount of available human attention becomes progressively less credible.
8. What happens if the agent is manipulated by something it reads?
Prompt injection becomes considerably more consequential when an AI system can act.
An ordinary model may read malicious instructions embedded in an email or webpage and produce a compromised answer. An agent with tools and permissions may allow the same manipulation to influence which actions it takes.
Microsoft explicitly warns that untrusted content from retrieved documents, tool outputs and other agents can enter the agent's reasoning loop and should not be treated as trusted instruction. Its current guidance recommends separating instructions from data, reapplying safety controls at agent boundaries and placing stronger gates around consequential actions.
OWASP's agentic AI work has similarly emphasized risks such as goal hijacking, tool misuse, identity and privilege abuse, memory poisoning and unexpected code execution.
For the CISO, the architectural question is whether untrusted information can quietly become authority.
The answer cannot rely entirely on the model recognizing malicious intent correctly every time. Permissions, action authorization, tool restrictions and intervention controls need to limit what a compromised reasoning process can actually accomplish.
9. What can the agent remember, and who can influence that memory?
Persistent memory is useful because agents can maintain context across tasks and sessions. It also creates another place where risk can survive.
Microsoft's shared-responsibility guidance identifies memory poisoning as a specific agent risk and recommends controls such as user and tenant isolation, validation of persistent information, provenance tracking and ordinary data-management practices around retention and access.
The security implications extend beyond prompt injection.
What personal or confidential information can the agent retain? How long does it remain? Can one user influence information later used for another? Is memory included in deletion and retention obligations? Can security teams determine where a consequential instruction, belief or assumption entered the agent's context?
These questions can feel less exciting than the agent demonstration.
They become more interesting when a supposedly resolved issue survives for another three weeks inside persistent memory.
10. What happens when one agent talks to another?
Multi-agent systems create another trust boundary because one agent's output can become another agent's instruction or context.
An orchestrator may divide work among agents with different identities, permissions and capabilities. The user may interact with only one visible interface while consequential activity happens several layers away.
That means problems can propagate, and accountability can become harder to reconstruct.
Microsoft recommends treating inter-agent messages as new trust boundaries rather than assuming information produced inside the agent environment is automatically trustworthy.
The organization should be able to determine which agents participated in consequential activity, which identities and permissions they used, and where authority changed hands as work moved through the system.
Our Human-Agent Accountability Guide describes this as preserving a human accountability anchor even when execution moves through an agent chain.
More agents should not mean less clarity about who ultimately owns the objective.
11. Can we stop the agent quickly enough?
A stop mechanism sounds almost embarrassingly basic until an autonomous system is moving through multiple connected tools.
Can the organization halt the current task? Can it revoke the agent's access? Does stopping the orchestration also stop credentials and downstream activity? Can security responders intervene without waiting for the business owner who created the workflow?
These are containment questions as much as AI questions.
Cybermaniacs' Human Override model treats override as more than a button. Effective Human Override requires an authorized person to be able to stop, reject, redirect, constrain, reverse or take control while doing so can still affect the outcome.
The related Intervention Authority question is whether the right people are actually permitted to exercise those powers.
CISOs should test that capability before deployment rather than assuming the platform's “disable” control means what everybody hopes it means.
12. What evidence will we have after the agent acts?
An agent saying “done” is not audit evidence.
Security and risk teams need enough traceability to understand what happened: which tools were invoked, which identity was used, which resources were affected, which inputs influenced the action and whether human approval or escalation occurred.
Microsoft recommends action-level audit logging for agent tool use, including inputs, outputs, identities and relevant decision information.
The amount of evidence required should scale with risk. Nobody wants to build an archaeology department for harmless internal agents, but accountability becomes difficult for consequential systems if the organization cannot reconstruct what happened.
With agents, evidence about actions matters more than confidence in the agent's summary of those actions.
That evidence also becomes part of Human-AI Risk Intelligence, because the organization needs to understand whether the technical and human control system behaved as intended after deployment.
13. Have we prepared the humans whose jobs are changing around the agent?
This question can disappear beneath identity, permissions and architecture because those issues are easier to specify technically.
Agent deployment changes human work.
Employees may move from performing tasks to supervising outcomes. Managers may become accountable for work they no longer directly observe, while teams may need to recognize escalation conditions, verify agent activity and intervene when authority needs to return to a person.
Those are new responsibilities rather than simply new software features.
Cybermaniacs' Agentic Readiness & Change (ARC) work focuses on that layer: changing roles, workflows, supervision, accountability, delegated authority and workforce capability as agents enter the operating environment.
Our broader AI Enablement & Change (AIECM) work addresses workforce readiness, confidence, behavior and adoption, while Human Risk Assessment & Assurance helps organizations understand whether the people relied upon inside AI controls possess the competency, judgment and organizational support required.
This is where security and change management become difficult to separate.
The safest technical agent can still sit inside a badly designed human system.
14. How will we know whether the human controls actually work?
This question deserves to be asked before deployment because the organization will otherwise discover the answer indirectly.
A workflow can contain human approvals, escalation, supervision and override without any of those controls performing particularly well. Reviewers may approve almost everything, supervisors may be overloaded, escalation may reach people without enough authority and override controls may exist technically while being too slow to prevent the relevant consequence.
Cybermaniacs uses Human Control Effectiveness to describe whether a human control relied upon in an AI-enabled process can actually perform the risk-reducing function expected from it.
Before deployment, the CISO should know what evidence will demonstrate that the control is functioning.
That might include intervention behavior, escalation outcomes, response times, scenario testing, supervisory workload or evidence that employees can recognize the situations the control expects them to catch.
The important part is deciding that evidence matters before the control becomes a permanent box in the governance diagram.
What should CISOs require before an AI agent goes live?
Different agents will need different controls, but a CISO should be able to get clear answers to a reasonably consistent set of questions.
The organization should understand what the agent can do, which identity and permissions it uses, what authority has been delegated and which actions require human involvement. It should know what happens when the agent leaves its intended mandate, how untrusted information is handled, what memory persists and whether other agents introduce additional trust boundaries.
Someone should remain meaningfully accountable, with enough visibility and authority to perform the role. Security teams should be able to contain the agent when necessary and reconstruct consequential activity afterward, while the workforce around the agent should understand the new supervisory and decision responsibilities it has been given.
The controls should scale with the authority.
An internal assistant with narrowly scoped read access creates a very different environment from an autonomous system able to modify production infrastructure, communicate externally or initiate financial activity. Applying the same governance burden to both would be inefficient; applying the same assumptions to both would be worse.
Agent readiness is bigger than agent security
The CISO has an essential role in agent deployment because agents create new identity, authorization, data, execution and monitoring risks.
Secure deployment still requires more than a technically secure agent.
It requires a clear relationship among the authority the agent has been given, the human who remains accountable for the outcome and the controls that determine when autonomous activity can continue.
Cybermaniacs uses Human-Agent Accountability to make that relationship explicit. Supervision Sufficiency helps determine whether the human supervisory capacity around the work is credible, while Human Control Effectiveness asks whether the human controls themselves actually reduce risk.
Human-AI Risk Intelligence provides the wider evidence layer for understanding whether human capability, behavior, psychology, culture, authority and working conditions support the controls the organization believes it has.
That is why the pre-deployment conversation should not end with “Is this agent secure?”
A stronger question is whether the organization is ready to let this agent, with these permissions, exercise this amount of authority inside this workflow—and whether the humans around it are prepared for what happens next.
That is the real deployment decision.