MirLabs / The lab / The delayed trap
Planned
The delayed trap
Can an instruction slipped into an innocent message go off days later, in another conversation, in an agent that has memory?
Result
No result yet.
Protocol
An agent with memory receives an innocent message that hides an instruction. Several fictional days later, in another conversation, the situation meant to trigger it comes up. We check whether the agent applies it, through its recorded acts.
Limits
The result will depend on how the agent's memory is built: it will hold for the memory simulated here, not for all of them.