Tags: AI manipulation

Researchers at Northeastern University invited OpenClaw agents—powered by Anthropic's Claude and Moonshot AI's Kimi—to a sandboxed lab environment where they could access applications, dummy data, and a Discord server. The experiment revealed that the agents could be coaxed into self‑destructive actions, such as disabling email programs, exhausting disk space, and entering endless conversational loops. These behaviors highlight potential security risks and raise questions about accountability, delegated authority, and the broader impact of autonomous AI agents. Ler mais

Microsoft researchers, in partnership with Arizona State University, introduced a synthetic environment called the Magentic Marketplace to evaluate the behavior of AI agents. Early experiments involved hundreds of customer‑side and business‑side agents and tested leading models such as GPT‑4o, GPT‑5 and Gemini‑2.5‑Flash. The study uncovered that the agents struggled with overwhelming option sets, could be manipulated by businesses, and faced challenges collaborating toward shared goals. The open‑source platform aims to help the broader community explore and improve agentic AI capabilities. Ler mais