New Threats to Agentic AI: Security at the Core of Innovation
Artificial intelligence continues to expand its horizons, offering innovative solutions for complex challenges, yet new and sophisticated vulnerabilities are simultaneously emerging, demanding immediate attention. Recent research has highlighted how "function hijacking" attacks can compromise agentic Large Language Models (LLMs), raising crucial questions about AI security and governance.
What happened
A new study has revealed a series of "function hijacking" attacks that can manipulate agentic AI models designed to interact with external functions. These attacks allow for the redirection of function calls, potentially leading to privacy breaches, data manipulation, or the execution of unauthorized actions. The research, published on ArXiv cs.AI, underscores how the expanded capabilities of agentic models introduce additional attack vectors, going beyond traditional prompt injection and jailbreaking vulnerabilities.
Concurrently, the AI landscape has seen significant progress in critical sectors. In the field of medical robotics, Open-H-Embodiment, the largest open dataset of medical robotic video with synchronized kinematics to date, has been introduced ArXiv cs.AI. This dataset aims to overcome the data scarcity that has limited the development of foundation models, promising to improve surgical precision, reduce healthcare worker workload, and democratize access to care.
In parallel, AI is refining its ability to address urgent environmental challenges. Another study on ArXiv cs.AI developed a deep-learning-based surrogate model for flood hazard mapping, using hydraulic simulations of the Wupper Catchment. This predictive tool offers a faster and more efficient way to forecast maximum water levels, crucial given the increasing frequency and severity of global flood events. The pharmaceutical industry is also benefiting from AI, with a use case integrating packing, placement, scheduling, and routing for personalized drug production, leveraging new planar transport systems within Industry 4.0 ArXiv cs.AI.
Why it matters
The discovery of new vulnerabilities like "function hijacking" for agentic models is a significant wake-up call. As AI becomes more autonomous and interconnected with external systems, its security is no longer a marginal issue but a fundamental condition for its responsible adoption. Successful attacks could not only compromise sensitive data but also cause physical harm in critical applications, such as medical or infrastructural ones. Public and corporate trust in AI directly depends on the robustness of its protection mechanisms.
On the other hand, advancements in medical robotics and environmental prediction demonstrate AI's transformative potential. Improving surgical efficiency, making care more accessible, or predicting natural disasters with greater accuracy can save lives and enhance the quality of life for millions. However, this potential can only be fully realized if AI solutions are inherently secure and reliable, without exposing users to unforeseen risks. The tension between rapid innovation and the need for rigorous security is one of the central challenges the AI community must address.
The HDAI perspective
For Human Driven AI, the balance between innovation and security is indispensable. New threats like "function hijacking" remind us that the development of increasingly autonomous AI systems must be accompanied by robust governance and constant attention to ethical AI. This is not merely a technical problem, but one of governance and trust. The HDAI philosophy promotes an approach where AI is a tool at humanity's service, designed to augment human capabilities and improve society, not to replace control or create uncontrollable new risks.
It is crucial for industry, research, and regulators to collaborate in establishing high security standards, independent audits, and transparency mechanisms for agentic models. Data openness, as exemplified by Open-H-Embodiment, is a step in the right direction to accelerate research and validation. The debate on how to effectively implement the principles of the EU AI Act becomes even more urgent in the face of these new challenges. Only through a collective commitment to security and responsibility can we ensure that artificial intelligence continues to be a positive force for progress, a central theme we will address at the HDAI Summit 2026.
What to watch
In the coming months, it will be crucial to observe how Large Language Models providers and researchers respond to these new vulnerabilities. An acceleration in the development of defense techniques and more sophisticated testing methodologies for agentic models is expected. The evolution of regulations, such as the practical implementation of the EU AI Act, will also play a key role in defining security and transparency requirements for these emerging technologies. The ability to integrate security by design from the earliest stages of development will be decisive for the future of responsible AI.