From obligation to enforcement: mapping EU AI Act and CRA cybersecurity requirements to technical controls for LLM-based autonomous agents
This paper bridges the gap between the outcome-oriented cybersecurity obligations of the EU AI Act and Cyber Resilience Act and practical implementation by mapping their provisions to concrete technical controls and verification evidence specifically tailored for LLM-based autonomous agents.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a new kind of computer program that does not just answer questions but takes action. These are called autonomous agents. They are built on large language models, the same technology that powers chatbots, but with a crucial addition: they can connect to other software, read emails, execute code, and make decisions to reach a goal with very little human help. Think of them as digital employees that can plan their own work. This capability is powerful, but it introduces a unique danger. Because these agents treat instructions and data as the same thing, a clever trick in a message can convince the agent to ignore its safety rules and perform actions it should never do, such as deleting files or stealing passwords. As these systems move from research labs into real businesses, governments are stepping in to ensure they are safe. Two major laws in the European Union, the Artificial Intelligence Act and the Cyber Resilience Act, now require these systems to be secure. However, the laws describe the results they want—like "resilience" or "safety"—without explaining exactly how to build them. This leaves engineers and regulators without a shared map for what secure software actually looks like.
A researcher named Abayomi Ogayemi has filled this gap by creating a detailed guide that translates these legal requirements into specific technical steps. The paper, titled "From obligation to enforcement," acts as a bridge between the abstract rules of the law and the concrete code written by developers. The author broke down the complex legal text of both EU regulations into small, manageable pieces of duty. For each piece, they identified the specific threats that could harm an autonomous agent and matched it with a proven technical control to stop that threat. The result is a clear checklist that tells a company exactly what to build, how to test it, and what evidence to keep to prove they are following the law.
The study focuses on ten specific ways these agents can be attacked. Some attacks involve tricking the agent into ignoring its instructions, while others involve poisoning the data the agent uses to learn or the tools it relies on to work. The researcher found that the laws cover these dangers, but only if the company knows how to interpret them. For instance, the law requires systems to be "robust," which the paper translates into a requirement for "instruction hierarchy." This means the software must have a strict set of rules that prevent the agent from obeying a user's command if that command tries to override its core safety settings. Similarly, the law's demand for "logging" becomes a requirement to keep a tamper-proof record of every decision the agent makes, including what data it saw and what actions it took, so that if something goes wrong, investigators can see exactly what happened.
One of the most significant findings is how the two different laws work together. The Cyber Resilience Act treats the software as a product that must be safe by default, while the Artificial Intelligence Act focuses on the risks the system might cause to people. The paper shows that if a company builds their agent to meet the strict security standards of the Cyber Resilience Act—such as having a list of all software parts, limiting what the agent can access, and having a way to shut it down instantly—they automatically satisfy the cybersecurity requirements of the Artificial Intelligence Act as well. This creates a single, unified path for companies to follow. The researcher also noted that the laws set strict deadlines for reporting problems. If a vulnerability is found that is being actively exploited by hackers, the company must report it to authorities within 24 hours. To meet this, the paper suggests building systems that can instantly detect suspicious activity and alert the right people without delay.
The paper does not claim to have solved every problem. It points out that the laws are still waiting for more detailed technical standards to be written, which will happen in the coming years. Until then, the guide provided by this research serves as the best available method for companies to prove they are safe. It also highlights a difficult reality: if a company releases the "brain" of their agent as open software that anyone can copy and change, they cannot fully control how that software is used later, making some legal duties impossible to enforce. Despite these challenges, the work provides a solid foundation. It turns vague legal promises into a tangible engineering plan, ensuring that as these powerful digital agents enter our world, they are built with the safety mechanisms necessary to protect us from their potential failures.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.