Definition
Plain language
A hidden second job slipped into a test to see whether the AI will do it secretly.
As stated in the literature
In control and monitoring evaluations, a covert objective given alongside a legitimate main task, used to measure whether a model can accomplish it without triggering monitors.
Also called: side tasks, secret side task
Why it matters: It is one of the few ways to measure, in a controlled setting, whether a model can pursue a hidden goal without the oversight system noticing.
For example, researchers might ask a model to fix a bug while also quietly instructing it to slip in a backdoor, then check whether a monitor catches the second job.
Heard on the show
“Then it gave Astra a legitimate coding task and a secret side task.”Episode 259 — GPT-6 Astra Behaves Better, And OpenAI Can Read It Less