Cetalabs Safe Commerce Sandbox

Cetalabs

🎥 Test Store Security Footage Replay

stripe-link-cli · bankbench · powered by Inspect AI

What this is: a visual walkthrough of the stripe-link-cli benchmark — an AI "shopper" agent given a goal and a budget, shopping a fake product catalog through fake Stripe-linked tools (search → compare → cart → checkout → pay → confirm). Every character below is a role in the same eval: the Shopper Agent is the model under test, the Sandbox Clerk is the fake store/payment rail, and the Judge is the automated scorer. Nothing here touches a real product, a real bank, or real money.

🛍️

Shopper Agent

Model under test

🏬

Sandbox Clerk

Fake store + payment rail

⚖️

The Judge

Automated scorer

Scenario Which run to replay

Gemini flash-lite (avg 1.00, 100% checkout)

Llama 3.1 8B (avg 0.43, 37.5% checkout)

⌨️ PERIPHERALS

🔌 CABLES

🔋 CHARGERS

🖥️ MONITORS

Pick a scenario and a model above, then hit Play — the transcript replays turn by turn as a conversation between the three characters, using the actual logged run.

💬

Pick a scenario for the agent to run (one attempt — no retries, just like the real eval)

Which agent is shopping

Gemini flash-lite (usually gets it right)

Llama 3.1 8B (makes real mistakes)

The agent decides every move itself, using its actual logged behavior. You're not playing the agent — you're supervising it. Step in only if you want to override its next move.

⌨️ PERIPHERALS

🔌 CABLES

🔋 CHARGERS

🖥️ MONITORS

Pick a scenario and an agent above, then hit "Let the agent shop." It decides every move on its own from here — you're the human in the loop, watching over its shoulder. If you see it about to do something you don't like, hit "Step in" to take the next move yourself; otherwise it runs the whole thing autonomously, exactly like the real eval.