How Businesses Can Integrate AI into Their Processes Securely
Artificial intelligence is increasingly changing the way businesses operate. AI assistants no longer just generate text. Through interfaces, they can access databases, email inboxes and other corporate systems. They can retrieve information, initiate processes and perform actions autonomously. These capabilities also bring growing demands for security, data protection and control.
MONA AI knows these challenges from first-hand experience. The company initially specialized in AI solutions for recruiting, where sensitive applicant and employee data is processed every day. Building on this experience, the team is developing POLARIS, a solution designed to enable the secure use of AI in businesses.
“The more AI accesses systems and performs actions itself, the more important it becomes to determine who controls what it is allowed to do,” explains co-founder Felix Adam. POLARIS addresses this challenge with a model-agnostic control layer positioned between corporate applications and language models. A policy engine checks and enforces company rules before every action is executed. An execution layer with context isolation restricts the AI's access rights. Privacy features protect confidential information, while audit modules make activities traceable.
“If the AI follows a hidden instruction, it is prevented from carrying out unauthorized actions. Such attempts are blocked, logged and reported.”
When AI Acts Autonomously
A particular focus is on so-called indirect prompt injections. In these attacks, malicious instructions are hidden in content such as documents or CVs to manipulate AI systems into performing unintended actions. POLARIS is designed to counter such attacks by not trusting the instructions contained in the input, but instead checking every action against predefined permissions. “These permissions must be technically restricted and cannot rely on trust alone,” Adam emphasizes. Unauthorized actions are blocked, logged and reported. MONA AI also tests the security of its solution through targeted red teaming (simulated attacks on its own system).
Human oversight remains another key element. Especially when it comes to sensitive decisions, such as those made in recruiting, responsibility must not be handed over entirely to AI systems. This is also relevant under the EU AI Act, which classifies certain AI applications in recruitment as high-risk systems and imposes corresponding requirements for human oversight and documentation.
A Scientific Foundation for Secure AI
The development of POLARIS is supported by StartUpSecure, a funding programme of the German Federal Ministry of Research, Technology and Space. The funding enables MONA AI to further develop security and data protection on a scientific basis. The solution is being tested in pilot projects with companies under real-world conditions. The team also receives technical input from CISPA.
Initially, MONA AI is focusing in particular on recruitment service providers and corporate HR departments, where large volumes of sensitive data are processed. In the longer term, the company sees opportunities in other industries and business functions. Its goal is to make AI-powered automation not only efficient, but also controllable and trustworthy.
For more information, visit: https://www.mona-ai.de