Columns follow the research loop; a risk sits at the step where the failure arises. Rows say who fails. Many are old failures that agents make faster and quieter. Click a risk to see the human failure that often lets it through (dark ring and dashed line) and the defences against it; click a defence to see what it covers; click empty space or press Esc to reset; arrow keys step through the risks. A ? marks an open question: nothing here convinces the author yet. Clicking a risk or a defence also shows an example you could adapt for your agent's instructions. A separate slide shows how an agent's error can pass the layers of checks: the Swiss cheese slide.
Examples, in plain words, of what to ask your agent. Adapt them; none has been tested. Most ask the agent to show you something, because a rule in a file does not prove it was followed (E4). Two habits help all of them: ask the agent to put its questions together, and to wait for your answer when it asks for your prediction, choice or brief.