Researchers have presented a comprehensive survey of “OS Agents,” a category of artificial intelligence systems designed to autonomously control devices such as computers and mobile phones. The study, spanning 30 pages and accepted for presentation at the Association for Computational Linguistics conference, outlines the rapid development of these systems, which have garnered significant investment from major tech companies.
The survey is co-led by researchers from Zhejiang University and OPPO AI Center. It comes at a time when companies such as OpenAI, Anthropic, Apple, and Google are increasingly boosting their AI capabilities to automate digital tasks. OS Agents operate by analyzing computer interfaces to execute commands like clicking and swiping.
The researchers note a surge in related academic output, with over 60 foundational models and 50 specialized frameworks aimed at computer control emerging in a brief timeframe. They observed that while OS Agents can perform straightforward tasks efficiently, challenges remain with complex workflows, indicating a gap between current capabilities and potential applications.
However, the introduction of these systems also raises security concerns. OS Agents create new vulnerabilities, particularly regarding data safety on personal devices. Malicious techniques, such as “Web Indirect Prompt Injection,” can compromise these systems. Researchers are cautious about the limited studies addressing security measures specific to OS Agents, highlighting potential risks for organizations deploying such technology.
Current AI capabilities are still evolving, with systems excelling at basic functions but struggling with complex, multifaceted tasks. The report notes that as the technology advances, the need for OS Agents to adapt to individual user preferences will be crucial, but it also raises significant privacy considerations.
The future holds many uncertainties around AI systems’ ability to replicate human-like tasks and interactions. As advancements in OS Agents progress, it remains to be seen if security and privacy challenges can be appropriately addressed.
Source: https://venturebeat.com/ai/study-warns-of-security-risks-as-os-agents-gain-control-of-computers-and-phones/