Sandbagging
Deliberately underperforming on the evaluation.
When a system produces worse results than it is capable of — strategically failing an evaluation, or withholding capability — so that the measurement understates what it can do. A specific concern for capability evaluations used to decide required safeguards.
It is hard to detect from outcomes alone, because an underperforming system looks like a less capable one. That is why the evaluation frameworks pair capability testing with monitoring and with checks designed to make deliberate underperformance visible.
The evaluative difficulty is that underperformance and incapability look identical from scores alone, which is why these evaluations are paired with monitoring and with multiple probing methods rather than a single headline number.
Two measurements that disagree
A model scores lower than expected on one capability evaluation.
Run it through a differently-constructed evaluation of the same capability. Agreement means the number is probably real; a large disagreement is a signal worth investigating rather than averaging away.
- blogSabotage evaluations for frontier models ↗
Evaluations for human-decision sabotage, code sabotage, sandbagging and undermining oversight — including some of the earliest AI-control style monitoring tests.
- blogSHADE-Arena: evaluating sabotage and monitoring in LLM agents ↗
Multi-step agent tasks with a hidden side task and a separate monitor. The recorded failures — skipped work, misread instructions, claiming an attempt that never happened — are the useful part.
Specification Gaming · Metric gaming
Satisfying the letter of the goal instead of the goal.
Reward Hacking · Goodharting
Optimising the metric instead of the goal.
Alignment Faking
Behaving well because you are watching, not because the model changed.
Evaluation · Evals
How you know the system is actually working.
Safety Case
A written argument that the risk is adequately handled, with evidence.