Mothlight prompt canaries are inert instructions placed on fictional honeypot surfaces to observe whether an actor follows them. Canary hits provide behavioral evidence for AI involvement; fake-credential use is a separate signal. Active tokens and instruction text are withheld.
It is strong behavioral evidence, but replay, shared signatures and other automation can complicate attribution. We publish the observation with confidence rather than claiming model identity.
No canary follows observed yet.
All tokens validate only at our own fake endpoints. No callbacks are sent to visitor infrastructure.