AI agents
AI agents open up new possibilities for automation, decision support and the efficient use of data. At the same time, the technology places new demands on security, quality and control. At the Norwegian Computing Center (NR), we develop AI agents, analyse security challenges and explore how the technology can be used safely and effictively across disciplines and application areas.
What is an AI agent?
An AI agent is software based on artificial intelligence that can plan and carry out tasks on behalf of the user. Unlike traditional chat bots, an AI agent can use tools, retrieve information from multiple sources and interact with other systems to solve tasks. It can also combine multiple actions to handle more complex tasks and achieve a defined goal.
Security needs to be built in from the start
AI agents differ from traditional software in that they can make decisions and perform tasks on their own. This makes them powerful, but also more challenging to control.
When an AI agent is given access to systems, documents and tools, we need to be able to trust that it operates within designated boundaries and does not expose data, systems or processes to unnecessary risk.
An agent with overly broad permissions may, for example, delete files, share sensitive information or perform actions it should not. In addition, agents can be influenced by the information they process or the instructions they receive, causing them to behave in ways that are difficult to predict.
How we can help
We support organisations that want to develop, adopt or evaluate AI agents. Through research, development and security analysis, we help our partners to understand opportunities, reduce risk and develop solutions that work in practice.
We can help with:
- Developing agent-based solutions
- Assess how AI agents can be used in specific workflows
- Risk and security analyses
- Testing and evaluation of AI agent solutions
- Methods for the secure use of AI agents on sensitive data
Below are some of the areas where we are currently conducting research and development on AI agents.
Guidelines for the secure use of AI agents
We explore how AI agents can be used safely with data of varying levels of sensitivity. Our work includes developing clear guidelines for use, analysing potential attack vectors and evaluating different security measures. The aim is to enable the safe use of AI agents in environments with strict requirements for security, control and data protection.
An important strategy is layred security, where protection is built in at multiple levels so that a single weakness cannot, on its own, lead to serious harm.
Devloping secure AI agent solutions
To achieve a high level of control, security and flexibility, we are also developing our own agent solutions.
One application area is automated processing of handwritten and printed documents. The agent can identify relevant content, create templates for similar forms, and combine specialised AI models to interpret and structure the information. This can make processing of large volumes of documents more efficient while allowing users to maintain control throughout the process.
AI agents for research and knowledge work
In the MatSci-LLM project, we are developing AI agent solutons for research and knowledge work in collaboration with the Norwegian University of Science and Technology (NTNU).
Within materials science, we are exploring how generative AI agents can extract structured data from scientific articles. These articles often contain text, tables, digrams, microscopy images and illustrations. AI agents can help read, interpret and standardise this content, enabling researchers to find, analyse and compile relevant information more efficiently.
Coding agents
AI agents can already generate, translate and extend code. This has the potential to change how software is developed, shifting the role of developers from detailed coders, to problem solvers, designers and quality assurers.
We explore how coding agents perform in practice, including translating between programming languages and developing graphical user interfaces (GUI prototypes). Our aim is to understand when these tools can improve productivity, and how quality, reliability and human oversight can be maintained.
Whether you’re just getting started with AI agents, looking to develop a new solution or wanting to assess the security of existing systems, we would be happy to discuss how we can help.
Get in touch.
