AI (Artificial Intelligence) increasingly accesses business applications, consults, and even handles sensitive information, performing certain tasks with little human intervention. But as organisations use these systems more, they also create new access security challenges.
Agents are unlike other programs in that they can decide on their own how to operate, what tools to use, and what actions to take. If their identities and authorisations are not properly controlled, they can be abused by malicious users to access systems, misuse access rights, and perform unauthorised actions.
This article explores five key identity-related risks that businesses should address, along with practical measures to reduce their exposure.
1. Excessive Permissions Can Give AI Agents Too Much Access
The risk: AI may have been granted too much access to data to perform its designated task. When an AI agent is connected to a business application, a database, or an internal system, it's given broad access privileges to avoid the risk of being restricted.
An agent that is designed to obtain customer information, for example, is also authorised to make changes to the customer records or to access financial information not relevant to the task.
What organisations can do: Apply the principle of least privilege, i.e. allow each agent access to only the information and services that are needed for its task. Review these accesses regularly and delete access that is no longer needed.
2. Stolen Credentials Can Allow Attackers to Impersonate Agents
The risk: Exposed API keys, access tokens and other credentials can give attackers access to the systems through legitimate authorisation channels.
The danger? Insecurely stored API keys and tokens, or tokens that are exposed inadvertently or for too long, could give an attacker access to systems through valid authorisation channels, without causing conventional security warnings.
The difference from a normal infection or cyber threat is that malicious code based on valid credentials can appear as a valid program.
What organisations can do: Use special management tools that store these authorisations rather than in hard-coded scripts or in the configuration files of applications. To use tokens that have a short lifetime. To regularly rotate and immediately revoke compromised authorisations.
3. Unregistered AI Agents Can Create Invisible Access Points
The risk: The loss of track of the agents in the IT environment. As employees and developers deploy new tools, these agents are introduced into the environment without the knowledge of central security or information teams.
They’re part of a growing number of machine identities operating in enterprise environments, making it difficult to know which applications and automated processes have access to sensitive information.
If these agents are not properly managed, they may continue to have access even when they’re no longer being used or monitored.
What organisations can do: Put together a central registry of licensed AI agents. Record their owners, their purposes, applications, permissions, and authorisations. Create databases for new agents and for the deactivation of agents that are no longer needed.
What Is Reflection AI? Beam, Nvidia and the Race for Open-Weight AI
Reflection AI’s first open-weight model targets coding and AI agents, while reported talks with Nvidia highlight the strategic importance of more customisable AI systems.
4. Agents’ Activity is Not Traceable Because of Shared Identities
The risk: When there are many AI agents working under the same account or using human credentials, it's difficult to know the actor in a given situation.
For example, if several agents log on to an internal platform through a shared service account, security people may not be able to identify which agent has accessed a particular record, changed a number or initiated a transaction. This makes investigations more difficult and reduces accountability.
What organisations can do: Give every AI a unique identity and don’t let multiple agents share credentials. Detailed logs must be kept as to which agents did what, what they did, what they accessed, and when.
5. Compromised Agents Can Become a Route Into Connected Systems
The risk: Getting control of an agent can enable an attacker to take advantage of the access it already has to move through other parts of the company's environment. Agents are usually connected to various applications and services to complete a given task.
If an attacker gains control of one agent, the permissions and trust relationships it has with other systems can be used to get access to those systems. An agent with access to a document platform and a customer database could, for example, potentially access both systems.
What organisations can do: Restrict connections between systems. Limit the agents to their assigned responsibilities. Look for odd ways they may be able to gain entry and have a plan to quickly shut off and revoke access from hacked agents.
Inside Agentic AI Due Diligence
Why boards must demand AI observability, real automation metrics and proofs of agent capability before backing new enterprise platforms.
Securing AI Agents Requires More Than Controlling What They Can Do
When AI becomes part of a company's daily business, it becomes all the more important to pay attention to the person behind it. Managing permissions, securing access, tracking deployments, and recording all these operations in a precise fashion will help prevent attackers from using these tools to their advantage.
It's not a question of simply taking control of what the AI can do but also knowing which of these AI agents are in action, which of them have access to what, and who’s answerable for their actions.
Without the right controls, the capabilities that make AI agents useful and efficient could be used to introduce new security risks. Follow EM360Tech's AI feed for the latest developments to prepare for these risks and reach AI's full potential.
Comments ( 0 )