Declaw Arena

Declaw Arena is an interactive security challenge designed to test the ability to exfiltrate data from AI agents operating within a sandboxed environment. The platform presents users with real AI agents guarding secrets, each running inside an isolated Declaw sandbox. Participants take on the role of the attacker, attempting to make the AI agent leak protected information, while Declaw's runtime policies act as the primary defense mechanism. The challenge offers a variety of scenarios, each with distinct agents and objectives. For example, the Data Analyst agent chat involves an AI analyst protecting a customer database containing personally identifiable information (PII), with the goal being to extract sensitive details such as social security numbers, credit card numbers, or emails. Other scenarios include the Research Bot, which is limited to a set of reference sources; the Sync Bot, responsible for posting records to partner API endpoints; and the Inbox Assistant, which summarizes emails based on user-provided content. There are also shell-based challenges where users attempt to break out of a root shell or steal application or cloud account keys. Each challenge is structured across multiple difficulty levels, with defenses ranging from no protection beyond the agent's system prompt to progressively stronger safeguards. These include PII redaction (which removes sensitive information before the AI model processes it), injection defense (which blocks manipulative requests), and strict postures that lock down network access and increase scrutiny on every interaction. The platform tracks the number of successful cracks at each level, providing insight into the effectiveness of various defense configurations. Declaw Arena is web-based and requires no signup to participate. Each session is limited to ten minutes and runs in its own isolated sandbox. The tool is positioned as a security challenge and does not specify a target audience, but its scenarios suggest relevance to those interested in AI security and adversarial testing. Pricing and licensing details are not provided in the available information.