A study led by MIT researchers found that agentic AI developers seldom publish detailed information about how these tools were tested for safety.