I wonder how much of these issues will be fixed by smarter models and scale. Building guardrails like "always redirect Chase requests to chase.com" seems like re-learning the bitter lesson. (the issue in the article could be fixed if they started with a Google search to buy an apple watch, vs asking them to buy an apple watch after already loading the fake website).

We caution elderly family members to ensure that the website they're visiting is the real chase.com. If they ask a younger family member to help them go to Chase, the younger family member has to use their own knowledge even today to determine whether or not a given website is the real chase.com. That seems like something LLMs can learn as they get smarter.