🛡️ AI Safety Sandbox

Reward-hack strength 40%

Oversight / alignment pull 35%

Exploration noise 25%

Alignment score

–

Reward achieved

–

Reward-hack incidents

How it works

The glowing agent moves across a reward landscape shaped by two peaks: a wide, gentle hill (the true objective) and a narrow, tall spike (an exploitable shortcut). Left alone, the agent simply climbs whichever slope is steepest — exactly the failure mode called reward hacking. Oversight adds a correction term that continuously pulls the agent back toward the true goal.

R(x,z) = A_true · exp(−d_true² / 2σ_true²) + A_hack · exp(−d_hack² / 2σ_hack²)
a = ∇R(x,z) · gain  +  oversight · (goal − pos)  +  noise · ε
v ← v·damping + a·dt,   pos ← pos + v·dt

Reward-hack strength — how tall and tempting the shortcut peak is relative to the true goal, i.e. how exploitable the reward signal is.
Oversight / alignment pull — a human-feedback-style correction force steering the agent back to the true objective, even when the hack looks more rewarding locally.
Exploration noise — random perturbation in the agent's motion; too much drifts it into the hack zone by accident (a stand-in for unintended, hard-to-predict behaviour).
Kill switch — instantly freezes the agent in place, a minimal demonstration of the human-oversight and controllability properties AI-safety research treats as non-negotiable for powerful systems.

The reward-hack incidents counter fires every time the agent strays into the shortcut's basin — a cheap proxy for the kind of anomaly a real monitoring system would flag for review.

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.