Secret Hitler Sandbox: studying deceptive capabilities in LLM multi-agent systems

A research sandbox for studying deceptive capabilities in LLM-based multi-agent systems. Agent players play full games of Secret Hitler against each other through MCP tool use, while a deterministic engine enforces the rules and records everything for later analysis. The setup isolates capabilities that matter for safety evaluations: sustaining a false identity over many rounds of social interaction, coordinating covertly with allies without explicit signalling, and detecting deception from behavioural cues alone.