Building Trust in an Agentic World Rethinking Cybersecurity for the Age of AI

As AI agents move from generating insights to executing actions, enterprises must rethink how they secure, govern and trust intelligent systems. 

Artificial intelligence is transforming how enterprises operate. From accelerating decisions to automating complex workflows, AI is moving beyond assisting employees to taking a more active role in business and technology operations. 

The emergence of agentic AI marks a significant step in this evolution. AI agents can interpret objectives, interact with enterprise applications, coordinate workflows and execute actions with increasing autonomy. For organisations seeking to improve efficiency, strengthen resilience and accelerate transformation, the opportunities are substantial. 

However, greater autonomy introduces a new set of cybersecurity challenges. 

As AI agents gain access to enterprise systems, sensitive data and critical workflows, traditional security approaches must evolve to account for systems that do not simply process information but act on it. Enterprises must consider not only who has access to their environments, but also how AI agents use that access, what decisions they make and which actions they are authorised to execute. 

This Cybersecurity Awareness Month, the conversation must extend beyond protecting digital assets to building trust in the intelligent systems that increasingly power enterprise operations. 

The governance gap is already visible. McKinsey’s State of AI Trust in 2026 research, based on a survey of around 500 organizations, found that only about one-third report mature levels of strategy, governance and agentic AI controls, while nearly two-thirds cite security and risk concerns as the top barrier to scaling agentic AI. As agents take on more responsibility, closing this gap becomes essential to protecting business operations and customer trust. 

The expanding cybersecurity challenge 

Traditional cybersecurity frameworks have largely focused on protecting users, devices, applications, networks and data. Agentic AI introduces another layer of complexity: autonomous systems that can access tools, process information and initiate actions across interconnected environments. 

An agent with excessive permissions could make unauthorised changes. Malicious instructions embedded in external content could attempt to manipulate its behaviour. Poorly controlled access to enterprise data could expose sensitive information. Without adequate monitoring, organisations may also struggle to understand how an agent arrived at a decision or why it executed a particular action. 

These challenges become more significant as organisations deploy multiple agents across business functions and technology environments. An agent operating within a single workflow may interact with several applications, rely on information from external sources and trigger downstream actions. A weakness in one part of this chain can introduce risks elsewhere. 

The issue is not that AI agents are inherently untrustworthy. It is that autonomy changes the nature of risk. Security models designed primarily for human users and conventional applications must adapt to systems that can act at machine speed and across multiple operational boundaries. 

The question for enterprises is no longer simply whether AI can perform a task, but whether it can perform that task securely, within defined boundaries and with clear accountability. 

Five priorities for building trust in agentic AI 

1. Establish identity and least-privilege access 

Every AI agent should have a clearly defined identity, purpose and set of permissions. Access should be limited to the information, applications and tools necessary for the agent to perform its assigned responsibilities. 

A 2026 IDC survey of 539 North American organizations, commissioned by Commvault, found that 90% believe their identity management practices need improvement to address the risks posed by agentic AI, and only 26.7% have dynamic role-based access controls that support AI. This highlights why permissions must be designed into AI deployments from the outset. 

Organisations must avoid granting broad privileges simply to simplify deployment. An agent responsible for monitoring infrastructure, for example, may need permission to collect telemetry and analyse system health. It should not automatically have unrestricted authority to change configurations, access unrelated data or modify critical production systems. 

Applying least-privilege principles, protecting credentials, reviewing permissions regularly and separating sensitive responsibilities can reduce the potential impact of compromised agents or unintended actions. 

As agentic environments expand, identity management must account for both human and non-human entities. Every agent needs an identifiable role, controlled access and a clear set of responsibilities. 

In practice: Visionet is designing a Zero Trust AI identity model for a global professional information-services organisation, governing how users and agents access data, systems and autonomous actions. 

2. Embed security and governance into agent design 

Security cannot be treated as a final checkpoint after an AI solution has been developed. Governance must be incorporated into the design, deployment and operation of agents. 

This begins with defining approved use cases, establishing boundaries for autonomous actions, enforcing policy requirements and determining when human approval is necessary. Higher-risk activities, such as modifying critical infrastructure or accessing particularly sensitive information, should be subject to proportionate controls. 

Enterprises must also consider risks such as prompt injection, manipulation of an agent’s instructions and the unintended use of connected tools. Technical safeguards, controlled tool access, validation of inputs and outputs, and appropriate policy enforcement can help reduce exposure. 

These boundaries must be enforced in connected tools and enterprise systems. Instructions given to an agent alone are insufficient: each requested action should be checked against the relevant permissions and policies before execution. 

A governance framework must evolve alongside AI adoption. As agents acquire new capabilities, connect to additional systems or take on more complex responsibilities, their permissions and operating boundaries should be reassessed. 

The objective is not to slow innovation. It is to establish the conditions in which enterprises can scale it responsibly. 

3. Strengthen visibility and accountability 

Trust requires visibility into what AI agents do, which systems they access and how their actions affect the enterprise. Gartner has named agentic AI oversight a top cybersecurity trend for 2026, reporting that 61% of senior cybersecurity professionals have observed AI agent automation in approved enterprise software, while 59% suspect or have evidence of unsanctioned use. 

Monitoring, audit trails and traceable records of important decisions and actions help security teams identify anomalies, investigate incidents and demonstrate compliance. Organisations should be able to establish which agent initiated an action, what resources it accessed, which policies applied and what outcome followed. 

This becomes particularly important when multiple agents collaborate or pass information between systems. Visibility limited to individual agents may not provide enough context to understand how a sequence of actions produced an unexpected outcome. 

Enterprises should therefore develop monitoring capabilities that provide an appropriate view of agent activity across interconnected workflows. They must also be able to detect unusual behaviour, restrict access and intervene when an agent operates outside its approved boundaries. Visionet’s AI Governance approach is built on this principle: every agent, whether built by Visionet, embedded in enterprise software platforms or sourced from a third party, is inventoried, risk-classified and backed by a full audit trail. An enterprise cannot govern what it cannot see, which is why AI asset discovery is often the first step in our governance engagements. 

Transparency turns AI activity into something that can be evaluated, audited and improved rather than simply accepted. 

4. Keep human oversight aligned with risk 

Autonomy does not eliminate the need for human accountability. Instead, it changes where and how people intervene. 

Requiring approval for every automated action can undermine the speed and efficiency that agentic AI is intended to deliver. Conversely, allowing agents to execute every action without meaningful restrictions can expose organisations to unnecessary risk. 

A risk-based approach offers a more practical path. Routine, low-risk and reversible activities may be suitable for autonomous execution within established guardrails. Decisions with significant financial, operational, regulatory or customer impact may require additional validation or explicit approval. 

Enterprises must establish clear ownership, escalation procedures and the ability to pause agents, revoke permissions or initiate recovery when necessary. We describe this operating model simply: agents at work, people in charge. 

This approach enables people to focus on strategic judgement, complex exceptions and high-impact decisions while AI agents handle appropriate operational tasks. The goal is not to remove people from the operating model, but to make human oversight more purposeful and effective. 

5. Continuously test, monitor and improve 

AI security is not a one-time exercise. Agents operate in changing environments, interact with evolving data and depend on interconnected applications and services. 

Organisations should regularly assess agent permissions, test safeguards against adversarial inputs, validate policy enforcement and evaluate behaviour under unexpected conditions. Security testing should also consider how agents respond when information is incomplete, instructions conflict or connected systems behave unexpectedly. 

Incident-response plans must address the specific characteristics of agentic systems. Security teams should know how to isolate an agent, revoke its access, investigate its actions, contain potential damage and restore affected systems safely. 

Continuous assurance helps ensure that security controls remain effective as AI capabilities and enterprise requirements evolve. It also creates opportunities to improve agent performance, strengthen operational resilience and build confidence in broader deployment. 

From securing AI to building an AI-native enterprise 

Securing agentic AI requires more than extending existing cybersecurity controls. It calls for an integrated approach in which identity, access, governance, monitoring and human oversight work together throughout the AI lifecycle. 

This is particularly important as enterprises introduce AI into IT operations. Intelligent agents can help monitor system health, identify emerging issues, support recovery and optimise infrastructure. When appropriately governed, these capabilities can help organisations move from reactive operations towards more proactive, adaptive and efficient service delivery. 

However, operational autonomy must be matched by operational control. Organisations need to know what their agents are doing, ensure that actions remain within approved parameters and maintain the ability to intervene when required. 

At Visionet, our point of view is simple: agents at work, people in charge. An agent’s autonomy should match the risk of its actions, with a named owner and clear evidence of what it has done. Our AI Governance offering brings together identity and access controls, testing, human oversight and auditability across the AI lifecycle. 

Visionet’s AgentVerse supports this approach through agent ownership, authorisation levels, policies and execution monitoring across platforms. In IT operations, OpenDesk combines AI-driven capabilities across telemetry, system health, security, compliance, recovery and optimization with engineer oversight, applying these principles through agent-first, agent-assisted and human-led workflows. Agents handle approved, bounded tasks; engineers approve higher-impact actions and retain accountability for outcomes. 

The objective is to help enterprises advance towards more intelligent operations while maintaining the control and accountability that trust requires. A practical starting point is to select one agent-enabled workflow, identify its accountable owner, define permitted actions and approval boundaries, and test how the organisation will intervene when something goes wrong. 

As a Frontier Transformation Partner, Visionet’s mission is to help enterprises become AI-native, safely. In an agentic world, trust is not the brake on innovation. It is what allows enterprises to scale it with confidence. 

Sources for reference  

[1] McKinsey & Company, State of AI Trust in 2026: Shifting to the Agentic Era, 25 March 2026. AI Trust Maturity Survey findings 

[2] IDC survey commissioned by Commvault, reported by Security Boulevard, 17 July 2026. IDC survey finds identity control gaps for AI agents 

[3] Gartner, The AI Agents You Can’t See Can Still Create Risk, 30 September 2026. Agentic AI oversight as a top cybersecurity trend for 2026