Agentic AI Security: Governance Gaps Organizations Can No Longer Ignore
The Next AI Security Challenge Is Already Inside Your Organization
Most organizations have spent some time developing an understanding of generative AI. They have introduced acceptable-use policies, established guidelines for handling sensitive information, and began evaluating the security and privacy implications of employees using AI assistants. These efforts remain important, but they address only part of the challenge ahead.
The next phase of enterprise AI adoption introduces a fundamentally different security consideration: agentic AI.
Unlike traditional generative AI tools that primarily respond to prompts, agentic AI systems can plan tasks, make decisions, interact with connected applications, and execute multi-step workflows with limited human involvement. An AI agent might schedule meetings, process invoices, modify customer records, write and deploy code, or communicate with vendors on behalf of an employee.
The distinction matters. Traditional AI generates information for a person to review and act upon. Agentic AI can act on that information, initiate additional actions, and continue executing tasks across business systems. This introduces new risks around identity, permissions, data access, accountability, and operational control.
For cybersecurity leaders, the concern is no longer limited to what information an AI model generates. It is what an agent can access, what decisions it can make, and what actions it can execute without human intervention.
flows with limited human involvement. An AI agent might schedule meetings, process invoices, modify customer records, write and deploy code, or communicate with vendors on behalf of an employee.
The distinction matters. Traditional AI generates information for a person to review and act upon. Agentic AI can act on that information, initiate additional actions, and continue executing tasks across business systems. This introduces new risks around identity, permissions, data access, accountability, and operational control.
For cybersecurity leaders, the concern is no longer limited to what information an AI model generates. It is what an agent can access, what decisions it can make, and what actions it can execute without human intervention.
Agentic AI Adoption Is Outpacing Governance
One of the most significant challenges is that organizations are actively looking to deploy AI functionality into products, services, or internal activities without building the essential governance to establish the foundation. Additionally, software-as-a-service (SaaS) providers are increasingly embedding AI agents and autonomous capabilities into platforms businesses already use. Collaboration tools, customer relationship management systems, developer environments, and productivity applications are gaining the ability to take actions rather than simply generate recommendations.
As these features become available through product updates, integrations, and subscription changes, organizations may find themselves running AI agents before security, IT, privacy, or compliance teams have evaluated them.
This creates a visibility problem. An organization may maintain an approved list of generative AI tools while having little understanding of the agents, integrations, and automated workflows operating across its environment.
Tevora’s 2027 CISO Trends Report: The AI Reckoning Is Here reinforces this concern. The report identifies AI governance as the top initiative for the coming year among surveyed cybersecurity leaders and highlights how AI adoption across employee tools, SaaS platforms, developers, vendors, and agents is outpacing organizations’ ability to maintain visibility and control. It also emphasizes the growing importance of identity, least privilege, data protection, and continuous security validation. Read the report.
This is an extension of the shadow IT problem, but with a significant difference: these systems can act on behalf of users and organizations. A productivity tool with access to internal documents presents one level of risk. An agent with the ability to send external communications, change financial records, or modify production systems presents another.
Organizations need to understand not only which AI tools are approved, but also what those tools are authorized to do.
flows with limited human involvement. An AI agent might schedule meetings, process invoices, modify customer records, write and deploy code, or communicate with vendors on behalf of an employee.
The distinction matters. Traditional AI generates information for a person to review and act upon. Agentic AI can act on that information, initiate additional actions, and continue executing tasks across business systems. This introduces new risks around identity, permissions, data access, accountability, and operational control.
For cybersecurity leaders, the concern is no longer limited to what information an AI model generates. It is what an agent can access, what decisions it can make, and what actions it can execute without human intervention.
Agentic AI Adoption Is Outpacing Governance
One of the most significant challenges is that organizations are actively looking to deploy AI functionality into products, services, or internal activities without building the essential governance to establish the foundation. do not always make a formal decision to adopt agentic AI. Additionally, Ssoftware-as-a-service (SaaS) providers are increasingly embedding AI agents and autonomous capabilities into platforms businesses already use. Collaboration tools, customer relationship management systems, developer environments, and productivity applications are gaining the ability to take actions rather than simply generate recommendations.
As these features become available through product updates, integrations, and subscription changes, organizations may find themselves running AI agents before security, IT, privacy, or compliance teams have evaluated them.
This creates a visibility problem. An organization may maintain an approved list of generative AI tools while having little understanding of the agents, integrations, and automated workflows operating across its environment.
Tevora’s 2027 CISO Trends Report: The AI Reckoning Is Here reinforces this concern. The report identifies AI governance as the top initiative for the coming year among surveyed cybersecurity leaders and highlights how AI adoption across employee tools, SaaS platforms, developers, vendors, and agents is outpacing organizations’ ability to maintain visibility and control. It also emphasizes the growing importance of identity, least privilege, data protection, and continuous security validation. Read the report.
This is an extension of the shadow IT problem, but with a significant difference: these systems can act on behalf of users and organizations. A productivity tool with access to internal documents presents one level of risk. An agent with the ability to send external communications, change financial records, or modify production systems presents another.
Organizations need to understand not only which AI tools are approved, but also what those tools are authorized to do.
A Real-World Warning: AI Agents and Banking Access
Recent reporting illustrates why agentic AI security is becoming an immediate concern.
The risks of poorly governed agentic AI are already becoming visible in financial services. In an October 22, 2026 report, Axios found that major U.S. banks, including Morgan Stanley, JPMorgan Chase, Bank of America, and Citi, were restricting third-party AI agent access after detecting agents operating within customer accounts. In some reported cases, agents appeared to take actions beyond users’ instructions, raising concerns about access to sensitive financial information, and the ability to initiate transactions. The issue highlights a critical distinction: a user granting an agent access to an account does not necessarily mean the agent should have permission to perform every action available within it. Organizations need controls that distinguish between viewing information, recommending actions, and executing transactions, with stronger authorization and human approval for high-impact activities. Source: AXIOS.
The important lesson is that authentication does not necessarily establish authorization for every action an agent may take.
An agent may have legitimate access to an account but still attempt an action the user did not intend. The same issue can occur within an enterprise. A finance agent may have access to payment records without needing authority to approve transactions. A development agent may need to modify code without having unrestricted access to production credentials. A customer service agent may need to retrieve account information without permission to change ownership or disclose another customer’s records.
The distinction between access and authority is fundamental to cybersecurity. As agents become more capable, organizations must establish enforceable boundaries around what they can observe, recommend, and execute.
Six Agentic AI Security and Governance Blind Spots
Most existing AI policies were designed around a human-centered interaction model: an employee submits a prompt, reviews the response, and decides what to do next. Agentic AI changes that model by allowing systems to chain decisions and execute actions without a person reviewing every step. The sprawl of AI and the ease in which agents can be created produce additional blindspots for any AI governance program. In Tevora’s experience, there are
Six areas that deserve immediate attention.
1. Identity and Access Management
AI agents do not fit neatly into conventional identity and access management (IAM) models. They are not human employees, but they may also operate differently from traditional service accounts. An agent might act on behalf of a user, use its own machine identity, inherit permissions through an integration, or access several systems through delegated authorization.
Without a clear identity, organizations cannot consistently review permissions, revoke access, or determine which actions an agent was authorized to perform.
Every deployed agent should have a traceable identity, a defined business purpose, a named owner, and documented permissions. Least-privilege principles should apply, with separate controls for reading information, modifying records, approving transactions, and executing code. Access should be reviewed regularly and revoked when the agent is no longer required.
2. Data Access and Information Protection
Agents can retrieve more information than a task requires, combine data from systems with different access restrictions, or transfer sensitive information between connected applications.
An employee may have legitimate access to several business systems without being authorized to combine all available information for every purpose. An agent that inherits broad permissions can create unintended exposure of customer records, financial information, proprietary documents, or intellectual property.
Organizations should evaluate how agents retrieve, process, retain, and transmit information. Data minimization, purpose restrictions, tenant isolation, output filtering, and controls on external tool calls should be incorporated into agent workflows.
The objective is to prevent authorized AI workflows from using otherwise accessible data in inappropriate ways.
3. Cybersecurity Threats and Connected Tool Abuse
Agentic AI expands the attack surface because external content can influence an agent’s subsequent actions.
Prompt injection is a particularly important concern. An agent processing an email, document, webpage, or support ticket may encounter malicious instructions designed to manipulate its behavior. If those instructions influence the agent, an attacker may be able to induce unauthorized tool use, extract sensitive information, or trigger unintended actions.
The potential impact depends heavily on the permissions available to the agent. An agent with limited read-only access presents a different risk from one that can send external emails, modify access permissions, execute code, or initiate financial transactions.
Security testing should therefore evaluate the complete workflow, including prompt injection resistance, tool authorization, output validation, credential handling, and safeguards against unintended actions. High-impact operations should have independent enforcement controls rather than relying exclusively on the AI model to recognize unsafe behavior.
4. Privacy and Regulatory Compliance
Agentic AI can complicate privacy compliance because workflows may collect, combine, infer, and act on personal information across multiple systems.
Under the General Data Protection Regulation (GDPR), organizations must consider lawful bases for processing, purpose limitation, data minimization, transparency, and appropriate security. Article 22 also addresses certain decisions based solely on automated processing that produce legal or similarly significant effects, subject to applicable exceptions and safeguards.
The California Consumer Privacy Act (CCPA), as amended, and its implementing regulations introduce additional considerations for certain personal information processing activities, including risk assessments and automated decision-making technology requirements in specified circumstances.
The EU AI Act establishes a risk-based framework for AI systems, with obligations depending on the system’s intended purpose, classification, and the organization’s role. Agentic capabilities do not automatically make a system high-risk, but they can materially affect how its potential impacts and safeguards should be evaluated.
Organizations should incorporate agent autonomy, connected systems, downstream actions, and human oversight into AI risk assessments and applicable data protection impact assessments (DPIAs). Compliance cannot be established through an acceptable use policy alone. It requires understanding how the system operates.
5. Accountability and Incident Response
When an agent performs multiple actions across connected systems, reconstructing the cause of an incident can be difficult without sufficient monitoring.
Organizations need visibility into agent identity, initiating requests, relevant instructions, data sources, tool calls, authorization decisions, approvals, and resulting actions. Logging should support investigations while respecting privacy and appropriate retention requirements.
Every agent should also have a named owner responsible for its approved purpose, operational performance, and ongoing risk management. Incident response procedures should establish how to suspend an agent, revoke credentials, stop connected workflows, preserve evidence, and recover affected systems.
Human oversight is only effective when the reviewer has sufficient context and the authority to intervene before a consequential action is completed.
6. Third-Party and Vendor Risk
AI agents embedded in third-party platforms introduce risks that extend beyond internally developed systems.
Existing vendor contracts may not clearly establish what an agent can do with company data, which sub processors it uses, how its actions are monitored, or who is responsible when an automated action creates a security or operational issue.
Organizations should extend vendor risk assessments to cover agent functionality, data access, permissions, security testing, logging, incident notification, model dependencies, and the ability to disable or restrict agent capabilities.
Contracts and procurement reviews should address these issues before deployment. New agent functionality introduced through existing SaaS providers should also trigger reassessment when it materially changes data access or operational authority.
Aligning Agentic AI Security with Regulatory and Industry Frameworks
Organizations do not need to create an entirely separate compliance program to address agentic AI. Existing frameworks provide a foundation, but they must be applied to the specific capabilities and risks of autonomous systems.
The NIST AI Risk Management Framework (AI RMF) offers a voluntary structure for identifying, evaluating, and managing AI risks through its Govern, Map, Measure, and Manage functions. It helps organizations establish accountability, document intended use, evaluate potential impacts, and prioritize mitigation.
ISO/IEC 42001:2023 provides requirements for establishing and continually improving an AI management system. It supports formal governance, defined responsibilities, operational processes, and ongoing oversight of AI activities.
The EU AI Act establishes legal obligations based on AI system classification, intended purpose, and organizational role. GDPR and CCPA requirements should also be considered where agents process personal information or influence decisions about individuals.
These frameworks serve different purposes. None independently guarantees that an agent is secure or that an organization is compliant. Technical testing, privacy analysis, documented governance, and ongoing monitoring must work together.
AI Risk Assessments and AIUC-1 Readiness: Turning Governance into Action
For organizations adopting agentic AI, an AI risk assessment provides a practical starting point for identifying weaknesses before they become operational or security incidents.
An effective assessment evaluates more than an AI policy or inventory. It examines the agent’s configuration, connected tools, identities, permissions, data sources, and ability to execute actions. It should also consider prompt injection, unauthorized tool use, data exposure, excessive autonomy, and the potential impact of incorrect decisions.
The result should be a prioritized view of control gaps, business risks, and remediation requirements. This gives leadership a practical roadmap for improving AI security rather than relying on assumptions about how an agent is supposed to behave.
AIUC-1 offers another useful reference point. Focused on AI agent security, safety, and reliability, the standard provides a structured basis for evaluating controls related to adversarial inputs, unauthorized actions, human oversight, reliability, governance, and third-party risk. For organizations unfamiliar with the framework, understanding its requirements, control domains, and relationship to existing standards is an important first step. Tevora’s guide, What Is AIUC-1? A Guide to the Newest AI Compliance Framework, provides a deeper look at how the standard works, what certification involves, and which organizations should consider pursuing it.
An AIUC-1 readiness assessment can help organizations evaluate their existing controls, identify gaps, and prepare evidence for a formal evaluation. Readiness is not the same as certification, and an assessment alone does not guarantee that an agent is secure. Its value is in establishing where improvements are needed and helping organizations demonstrate that appropriate controls are operating.
These services can complement NIST AI RMF alignment, ISO/IEC 42001 implementation, privacy assessments, and broader cybersecurity programs. Together, they help organizations move from high-level AI policies to measurable security and governance practices.
What Organizations Should Do Now
The immediate priority is visibility and accountability. Organizations should inventory every AI agent in use, including capabilities embedded in SaaS platforms, developer tools, and workflow automation products. Each agent should have a documented purpose, owner, permissions, integrations, and risk classification appropriate to its use.
Leaders should establish agent identity and least-privilege standards, update AI policies and threat models, and incorporate autonomous actions into privacy and compliance assessments. Vendor contracts should be reviewed to address agent access, data handling, monitoring, incident reporting, and responsibility for failures.
Finally, every deployed agent should have a named owner and a process for periodic access reviews, testing, monitoring, and retirement. AI risk assessments and AIUC-1 readiness evaluations can help identify gaps and prioritize the controls that matter most.
In Conclusion: Govern AI Agents Before They Become an Incident
Agentic AI creates meaningful opportunities to improve productivity and automate complex business processes. However, its ability to plan, access tools, and execute actions introduces security and governance requirements that traditional AI acceptable-use policies were not designed to address.
Organizations need visibility into deployed agents, controlled identities and permissions, protection against manipulation and data exposure, reliable monitoring, clear accountability, and vendor agreements that reflect autonomous capabilities.
The organizations best positioned to adopt agentic AI safely will be those that treat governance as an operational capability. That means assessing risk before deployment, validating controls through testing, monitoring behavior in production, and revisiting decisions as agent capabilities evolve.
Organizations don’t need to slow agents down. They need to treat agents as in scope now and build the controls that let them run safely.
How Tevora Can Help
Tevora helps organizations build the governance foundation for AI assurance through AI risk assessments, ISO 42001 readiness, and third-party risk programs adapted for agentic systems. Tevora also helps with the design, optimization, or operation of key AI program activities. Establishing that foundation before a customer requirement or deadline arrives makes readiness easier to manage.
To discuss where your AI program stands, call (833) 292-1609 or email [email protected].

