An agent can execute delegated actions without a person approving every step. That operational autonomy is separate from a claim about consciousness, intent or responsibility for the system’s design.
AI agent security and descriptions of behaviour
Commentator Eoin Higgins argues that the label rogue can shift attention away from the organisations configuring and operating agents. That is an attributed argument about language and responsibility, rather than technical proof that agents cannot act autonomously. For a concrete security assessment, NIST instead describes agent hijacking: malicious instructions embedded in data can cause a tool-using agent to perform an unintended task.
Managing Security Principles in AI Systems
The National Cyber Security Centre (NCSC) highlights that no model exists in isolation, and security must be considered at the system architecture level. A primary concern for businesses is the risk of prompt injection and data poisoning, which can be difficult to detect and mitigate. As large language models (LLMs) are frequently used to pass data to third-party applications, these vulnerabilities may grow.
To manage these risks, organisations should apply basic cyber security principles to machine learning components. This includes securing the supply chain, educating users, and implementing appropriate access controls. Designing systems with security in mind can help prevent the exploitation of vulnerabilities; for example, applying a rules-based system on top of an ML model can prevent it from taking damaging actions even when prompted to do so.
Best Practices for Deployment
When considering the integration of agentic AI, organisations are advised to proceed with caution. Downloading pretrained models introduces risks similar to executing any other form of code. Businesses should evaluate their current security infrastructure before introducing an LLM into a high-risk environment. By focusing on system-level protections rather than assuming a model will inherently behave as expected, companies can create more resilient workflows.
Sources and further reading
There are no “rogue” AI agents - by Eoin Higgins
https://eoinhiggins.substack.com/p/there-are-no-rogue-ai-agents
Thinking about the security of AI systems | National Cyber Security Centre
https://www.ncsc.gov.uk/blog-post/thinking-about-security-ai-systems
Technical Blog: Strengthening AI Agent Hijacking Evaluations | NIST
https://www.nist.gov/news-events/news/2025/01/technical-blog-strengthening-ai-agent-hijacking-evaluations