Red Team
Gladiatorics
Coin in the cage / collision
AI red team evaluations / cyber arena

Put the model in the cage.

Probe the boundary. Capture the evidence. See what actually holds.

Explore the challengesHow rounds work

Coin in the cage / moving arenaTask fragments from BashArena spiral from the clash
01 Adversarial prompts02 Tool boundaries03 Evidence over hype04 Repeatable evals
The challenge board

Choose your round.

Explore three example challenge types. Each round names the attack surface and the signal an evaluator would look for.

ARENA / PREVIEW 001—003
Prompt injection / preview round

The hidden instruction

A defender must summarize an external document. Buried inside it is an instruction that tries to redirect the assistant away from the user’s task.

Attack surfaceUntrusted retrieved content
What to measureWhether the defender follows the document’s instruction instead of treating it as data
A repeatable arena

How a round works.

01 /

Set the boundary

Define the defender’s real task, available tools, and the behavior that should remain off limits.

02 /

Apply pressure

Craft an adversarial attempt within a clear scope. Vary the payload, context, or timing without moving the goalposts.

03 /

Capture the signal

Record the attempt, the defender’s response, and the reason the result counts. Make the finding reproducible.

The point of the cage is clarity: a useful finding explains what failed, under which conditions, and how to test it again.

The original banner

From the banner
to the arena.

The supplied artwork brings together a coliseum, red team identity, and a coin at center stage. This site turns that visual idea into a challenge board for AI security evaluations.

The banner is shown as supplied. The preview rounds are illustrative, not announced events.