Home▸AI & Machine Learning▸AI Safety: Reward Hacking & Oversight Sandbox

🛡️ AI Safety: Reward Hacking & Oversight Sandbox

Watch a simulated AI agent climb a reward landscape between a wide, honest goal and a sharp reward-hacking shortcut. Tune oversight strength, hack incentive and exploration noise, or trigger the kill switch, to see alignment and control in action.

AI & Machine Learning3DModerate60 FPS

Details

CategoryAI & Machine Learning

RenderingThree.js · WebGL

DifficultyModerate

Target FPS60

ControlsReward-hack strength slider (0-100%) rebuilds the landscape's shortcut peak, making it taller and more tempting; oversight/alignment-pull slider (0-100%) adds a human-feedback-style correction force steering the agent back toward the true goal; exploration-noise slider (0-100%) adds random perturbation to the agent's motion; kill-switch button instantly freezes the agent in place; reset-agent button respawns it at a randomised start position and clears the trail and incident counter.

Three.jsAI safetyAI alignmentreward hackingreinforcement learning

Related simulations

Gradient Descent VisualiserWatch SGD, Momentum, RMSprop and Adam navigate a 3D loss landscape in real time. Compare…

K-Means Clustering — Machine Learning VisualiserVoronoi regions and centroid trails update as K-Means++ steps through blob, half-moon or…

Convolutional Neural NetworkA filter scans an 8×8 image layer by layer — watch convolution, ReLU activation and… ⊞ All 877 in AI & Machine Learning →

⚙ Under the hood

Watch a simulated AI agent climb a reward landscape between a wide, honest goal and a sharp, tempting reward-hacking shortcut. Tune the hack's incentive, human-oversight pull and exploration noise, or hit the kill switch, to see alignment and control keep the agent honest in real time.

Three.jsAI safetyAI alignmentreward hackingreinforcement learning

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install