Cyberattack? Help24/7

AI Security & Risks · Research commentary

Access Rights and Permissions for AI Agents

Published: Last content update: 6 min read

When an employee is given access to a business system, it is usually determined which information that person may view and which actions they may perform. For AI agents, this demarcation is at least as important. These systems can independently retrieve information, process instructions and, through integrations, make changes in business applications. Yet when AI agents are set up, their access rights are not always sufficiently limited to their actual tasks. An agent intended to answer customer questions may therefore gain access to data or functions that are not necessary for that work. The risk arises when a system not only has access to information, but also has the authority to perform actions that fall outside its original assignment.

When access rights become a security risk

AI agents differ from traditional chatbots in that they can take actions through external tools and system integrations. An agent may, for example, be connected to a CRM system, an inventory management system, an email environment or financial software. Which actions are possible is determined in part by the rights of the accounts and API connections through which the agent gains access. When an integration is intended only to retrieve order information but also allows data to be modified or deleted, the agent has more permissions than necessary.

In cybersecurity, this is referred to as Excessive Agency: a situation in which an AI system is given too much functionality, too many permissions or too much autonomy. The Open Worldwide Application Security Project (OWASP) describes this as a distinct security risk category. The danger is not limited to malicious attacks. Misinterpreted instructions, incorrect assumptions or unexpected system behaviour can also lead an AI agent to perform actions that were not anticipated. The consequences depend on the systems to which the agent is connected and the rights available within them.

Consider a wholesaler that uses an AI agent to handle questions about orders, delivery times and returns. To answer these questions, the system needs access to certain customer and order data. If the same agent can also independently modify invoice details, process refunds or cancel orders, however, the risk profile changes considerably. A misinterpretation of a customer request can then feed directly into the company's records. Where customer rights are insufficiently separated, there is also a risk that one client's information becomes accessible to another user. Security must therefore determine not only which applications an agent can reach, but also which specific data and functions within those applications may be available.

An important starting point for limiting these risks is the least privilege principle. This means that a system is granted only the minimum permissions required to carry out a particular task. An agent that needs to consult information, for example, does not need the right to delete records. When an action temporarily requires broader rights, those rights can be granted for a limited period and under predefined conditions. This prevents an AI agent from permanently holding permissions that are only occasionally needed.

Technical boundaries and human oversight

Setting instructions in an AI model is not in itself sufficient to prevent unwanted actions. A system instruction stating that an agent may not delete data does not offer the same protection as a technical access restriction that actually makes deletion impossible. AI models can misinterpret instructions or be influenced by information they process during their work. Access controls must therefore function independently of the language model. The connected application or a separate authorisation layer must check, for every relevant action, whether the agent is authorised to perform it.

The identity under which an agent operates also deserves attention. When multiple AI agents use the same administrator account or the same general access key, it becomes more difficult to distinguish responsibilities and to restrict permissions individually. A separate digital identity for each agent makes it possible to assign access rights specifically, record activities and revoke access when necessary. In its guidance on identity management for AI agents, the US National Institute of Standards and Technology (NIST) emphasises the importance of separate identities, temporary authorisations and auditable decision-making on access rights.

For actions with potential financial, legal or operational consequences, additional human approval may be necessary. Examples include changing payment details, approving financial transactions or deleting important business information. The approval procedure must be technically tied to the specific action, so that a general consent cannot be used to carry out other actions. Human oversight does not mean that every simple action must be reviewed manually. The aim is an appropriate balance between autonomous execution of limited tasks and explicit authorisation for actions with a higher risk profile.

In addition to limiting permissions in advance, oversight during use is necessary. Organisations must be able to establish which agent performed an action, when it took place and on behalf of which user or process. Logging and monitoring make anomalous behaviour visible and support the investigation of any incidents. These records must be protected independently of the agent, so that the system cannot alter or erase its own activities unnoticed. It must also be possible to block access immediately when unexpected behaviour is detected.

Access rights also warrant periodic review. Over its lifetime, an AI agent may acquire new functions, data sources or system integrations. As a result, permissions can gradually expand without the original risk analysis being repeated. This phenomenon, also known as scope creep, increases the likelihood that an agent eventually has access to more systems than it was originally designed for. Particularly with AI agents that support multiple business processes or work together with other agents, oversight of their combined permissions can be lost.

Conclusion

The security of AI agents is not determined solely by the reliability of the AI model used. Access rights, connected applications and permitted actions largely determine the consequences that unexpected or unwanted behaviour can have. As organisations give AI agents more autonomy, it becomes necessary to distinguish between what a system can technically do and what it is actually permitted to do.

Responsible implementation therefore requires precisely defined permissions, independent authorisation checks, appropriate human approval and continuous monitoring of system activity. These measures not only limit the consequences of potential attacks, but also protect organisations against operational errors within their own automated processes. The central security question is therefore not how many tasks an AI agent can perform autonomously, but under what conditions that autonomy can responsibly be permitted.


Sources

1. OWASP – LLM06:2025 Excessive Agency
Description of security risks arising from excessive functionality, permissions and autonomy in AI systems, including preventive measures.
https://genai.owasp.org/llmrisk/llm062025-excessive-agency/

2. OWASP – AI Agent Security Cheat Sheet
Technical recommendations on access restrictions, authorisation, human approval and the security of AI agents.
https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html

3. NIST – Design Tenets for Agent Identity Management Systems
Principles for digital identities, least privilege, temporary authorisations and auditability of AI agents.
https://www.nist.gov/blogs/cybersecurity-insights/design-tenets-agent-identity-management-systems

4. Canadian Centre for Cyber Security – Careful Adoption of Agentic AI
Guidance on the secure adoption of AI agents, including risk assessment, access management and operational security.
https://www.cyber.gc.ca/en/guidance/careful-adoption-agentic-ai