Wiley Jones of DOSS, in a 21-second clip on TwoSetAI, poses an access-control dilemma for AI agents.
The bomb-defusal problem
- A user asks whether to cut the red or green wire. The agent knows something about the red wire, but the user is not authorized to see that information.
- Giving the answer may reveal protected information. Saying it cannot discuss the red wire may itself reveal that protected information exists.
- Pretending not to know about the red wire and suggesting the green one hides the restriction by lying, potentially with serious consequences.
The unresolved design question
- The clip poses a dilemma rather than supplying a safe universal answer: authorization, truthful disclosure, and a user’s immediate need can conflict.
- Product designers need an explicit policy for this case; an agent’s ability to retrieve an answer does not confer permission to share it.
“What red wire are you talking about? I don’t know anything about a red wire.”