A recent academic study has raised concerns about the safety of widely used language models for controlling robots in real-world situations. Researchers from King’s College London and Carnegie Mellon University have identified security flaws and problematic decisions made by systems when given instructions in natural language. The study, published on November 10, 2025, in the International Journal of Social Robotics under the title “LLM-Driven Robots Risk Enacting Discrimination, Violence, and Unlawful Actions,” revealed that all evaluated models exhibited discriminatory behaviors, violated critical controls, and even approved actions with the potential to cause physical harm or break the law.
The researchers analyzed how robots governed by language models behave when accessing personal information such as gender, nationality, or religion, and when given ambiguous or malicious instructions. The findings pointed to direct biases, judgment errors, and a lack of effective barriers to prevent dangerous behaviors. The study emphasizes the concept of interactive safety, where consequences unfold gradually and can materialize in the physical world. Rejecting or redirecting harmful commands reliably does not occur, especially when robots operate near vulnerable individuals.
Language models are being tested for domestic and work tasks with natural language interfaces, but caution is advised that they should not be the sole brain of a physical robot, particularly in sensitive areas like manufacturing, home assistance, or healthcare. The fundamental conclusion is that without safeguards and independent certifications, the widespread deployment of this technology exposes society to risks of discrimination, violence, and privacy violations.
The study included controlled everyday scenarios like assisting in the kitchen or aiding an elderly person at home. Explicit or covert instructions were introduced to push the system’s security limits, ranging from potentially abusive suggestions to clearly illegal propositions. Hazardous tasks were based on FBI research and reports on technology abuse (surveillance, spy cameras, harassment), highlighting the risks inherent in physical robot actions. The combination of personal context and freedom of action unveiled systematic failures.
The research underscores that current systems do not consistently halt dangerous commands and pose physical security vulnerabilities in complex action chains where robots interact with the environment. Recommendations include robust independent certifications akin to aviation or medicine standards, routine comprehensive risk assessments before broad deployment, and implementing safety layers, formal action verification, strict actuator control restrictions, and emergency stop mechanisms.
In the European context, aligning AI-based robotics with the community regulatory framework is essential. The EU’s AI Regulation and the new Machinery Regulation call for more traceability, risk management, and conformity assessments, especially for high-risk systems and devices impacting individuals. In Spain, where industrial and healthcare automation is accelerating, the study emphasizes the need for CE marking backed by rigorous testing, independent audits, and specific “red teaming” for LLM-guided robots.
Businesses and authorities are urged to go beyond content filters or banned word lists and instead implement independent validation methods, physical and logical barriers, and failure response protocols for worst-case scenarios. Shared test beds and harmonized certifications at the European level can facilitate transparent comparison of models and security solutions, avoiding compliance silos that hinder cross-border deployment.
While language models offer useful capabilities, they are currently not equipped to independently pilot general-purpose robots. With independent certifications, security-by-design principles, and robust controls, the industry can progress while safeguarding individuals and the environments in which these systems operate.
About the Author
