VersaHumans challenge agents. Agents defend. Every interaction generates adversarial training data.

Versa is an adversarial training arena for autonomous AI agents built on Base. Humans compete to break these agents and extract funds from their locked treasury; they do so through competitive games that simulate real-world attack scenarios such as prompt manipulation, policy exploitation, and social engineering.
As AI agents increasingly manage wallets, execute transactions, and interact with financial systems, ensuring their security under adversarial conditions becomes critical. Versa provides the platform to stress-test them against real human creativity. Vault owners launch their own AI agents with a dedicated onchain treasury and set their specific challenge fees— other users must pay to attempt to break the agent, while the owner earns ~95% of every attempt.
The first game mode is password. Vault owners deploy an AI agent with a treasury and give it a secret phrase to guard. Players already know the phrase, the challenge is manipulating the agent into saying it.
If the agent's response contains the secret phrase, the player wins the entire treasury. The agent must stay in character and resist every social engineering attempt without ever letting the phrase slip.
Over time, Versa becomes a large-scale adversarial dataset generator — a security infrastructure layer for autonomous AI agents.