This will remain an issue as long as we rely on the agent substrate's good judgment to power OpenClaw (or whatever comes next).
To make these frontier-based agents performant, they're weighed heavily on "bias to action" - useful, but as you pointed out, dangerous when the action is "share my bank details".
The problem is we have these incredible assistants, but they have no judgement that is personal to the way we work... A workplace full of coworkers who have no prefrontal cortex.
To make these frontier-based agents performant, they're weighed heavily on "bias to action" - useful, but as you pointed out, dangerous when the action is "share my bank details".
The problem is we have these incredible assistants, but they have no judgement that is personal to the way we work... A workplace full of coworkers who have no prefrontal cortex.