collusion

In AI control, coordinated behavior by multiple untrusted AI instances that helps them subvert a control protocol, such as an actor attacking only when a shared signal is present while an untrusted monitor suppresses the corresponding alarm.

Metadata

aliases
en
  • AI collusion
  • actor-monitor collusion
  • model collusion
  • policy-monitor collusion
concept_id
collusion
preferred
en
collusion
relations
relations
prerequisite
related
status
active

Documents: about (1)

Documents: mentions (1)

prerequisite (1)

related (1)