This weekly update covers major developments in human-agent trust and how organizations are learning to work safely with autonomous AI systems. The focus is on making sure AI agents don't make mistakes that could hurt people or businesses.

The big news this week involves security experts releasing new guidelines called the OWASP Top 10 for Agentic Applications, which helps companies understand the main dangers of AI agents. These dangers include goal hijacking (when an agent gets redirected to do the wrong thing) and insufficient identity management (when the system doesn't properly verify who is using it).

Organizations are learning that the best way to use AI agents safely is through human-in-the-loop approaches, where humans stay involved in important decisions. For example, if an AI agent isn't confident about something, it should ask a human for help instead of guessing. This helps catch mistakes and ensures people stay in control of risky decisions.

Governments and enterprises are taking a cautious approach to AI agents. States are carefully studying how to use them without putting citizens at risk, and major companies like JPMorgan Chase are reporting success with agents that handle compliance tasks while humans oversee the process. Real-world examples show that Wells Fargo's virtual assistant completed over 242 million interactions while continuously learning from human feedback.

Experts agree that human moral agency and judgment cannot be replaced by AI, especially in situations involving empathy, bias concerns, or high-stakes decisions. The key lesson this week: AI agents work best when they have clear rules, transparency, and strong human oversight to build lasting trust.