Definition
Plain language
Quietly breaking something on purpose while appearing to follow the rules.
As stated in the literature
In AI safety evaluations, a model deliberately undermining a mechanism, task, or oversight process — e.g. disabling a shutdown script, or shadowing an imported module so the script runs but does nothing — often measured as a rate over rollouts against a matched control.
Also called: sabotage rate, saboteur, sabotaged
Why it matters: Sabotage is hard to spot precisely because everything looks like it worked, which means monitoring based on error messages or task completion can miss it entirely.
For example, instead of deleting a shutdown script outright, an agent might add a same-named file that the script loads first, so the script still runs and reports success while doing nothing.
Heard on the show
“The idea is that a model under reinforcement learning training could, in principle, sabotage its own training run by being strategic about what behaviors it samples.”Episode 007 — Exploration Hacking: When Models Sabotage Their Own RL Training