Add safety constraints to AI agents before allowing external actions.

Give AI agents a limited action list and require manual approval for changes like purchases, account edits, or public posts, as seen in the rogue gym agent case.