Artificial intelligence is now embedded in everyday business and consumer technology, from generative AI and large language models (LLMs) to computer vision, spatial sensing, human-machine interaction and combined "multimodal" systems. As organisations adopt these technologies, they also inherit a new and fast-growing class of risks. AI models can be manipulated through data poisoning, evasion and backdoor or Trojan attacks; they can leak private or proprietary data; and they can produce biased, unsafe or simply incorrect outputs, including "hallucinations" in generative systems. These weaknesses are hard to detect with conventional cybersecurity tools, and they become harder still when AI is connected to the physical world through cameras, sensors and edge devices, where subtle manipulations can change a system's behaviour without being noticed.
Two groups feel this gap most acutely. Companies that build AI models lack mature, standardised ways to test their systems for security and trustworthiness, and doing so manually is slow and costly. Companies that adopt AI often do not have the in-house expertise to evaluate or protect the models they rely on. At the same time, regulators and society are raising the bar: frameworks such as the EU AI Act, together with the EU's Ethics Guidelines for Trustworthy AI and the wider European AI strategy, expect AI to be demonstrably safe, robust and fair, while public trust in AI across Europe remains low.
DeepKeep addresses this need by developing an AI-native security and trustworthiness platform that protects AI applications across LLMs, vision, spatial sensing, human-machine interaction and multimodal models. Unlike approaches that focus on language models alone, the platform is designed to cover the full range of AI systems, including the visual, sensing and physical dimensions that are essential for real-world deployment. Its aim is to provide holistic protection across the entire AI lifecycle, from development and testing through to runtime operation.
The project pursues a set of connected objectives: to automatically discover and analyse the weak points of an AI model; to detect and mitigate threats specific to physical and multimodal systems; to deliver trust, security and privacy "guardrails" and a runtime AI firewall for generative and multimodal AI; to adapt these protections for constrained edge devices; to support the evaluation of, and compliance with, emerging AI regulation; and to automate and scale the platform so it can be deployed flexibly across cloud, on-premise, multi-cloud and disconnected environments. Sustained communication, dissemination and exploitation activities accompany the technical work to bring the results to market and to the wider AI community.
The project's pathway to impact runs from a working system prototype demonstrated in operational environments, reached in the first half of the action, towards a fully qualified system operating in real conditions by the project's end. By making rigorous AI security and trustworthiness testing automated, affordable and broadly accessible, DeepKeep aims to help organisations adopt AI safely and responsibly. The expected impact is significant in scale: the AI Trust, Risk and Security Management (AI TRiSM) market that defines DeepKeep's domain is projected to grow several-fold over the second half of this decade, with Europe among its fastest-growing regions. Beyond its commercial potential, the project supports the EU's ambition to be a global reference point for trustworthy AI, strengthening confidence among citizens and businesses, reinforcing European capability in AI security, and contributing to responsible innovation and skilled employment in a strategically important field.
