Design of Chimera

Overview of Chimera

Here, we provide a detailed description of the Chimera framework, covering its overall architecture, agent society construction, and threat simulation mechanisms. To ensure transparency and reproducibility, we also release all the prompts used in Chimera. These prompts were not arbitrarily designed; rather, they were iteratively refined based on feedback and insights from both security and AI experts. Through multiple rounds of preliminary experimentation and consolidation, we distilled a set of robust and effective prompts that faithfully capture realistic enterprise operations and insider threat behaviors.

Organization Profiling

Chimera begins by modeling the organization’s structure (employees, roles, and objectives) and system settings (OS, applications, services). These factors directly shape both application- and system-level logs. To support different needs, Chimera can either automatically generate realistic organizational profiles or accept user-defined ones, ensuring flexible yet faithful simulations.

Agent Society Construction

Once the organizational setup is ready, Chimera builds an LLM-based agent society where each employee is represented by an agent bundle. Each bundle combines multiple LLM agents with specialized tools (e.g., terminal, browser, file operations) to mimic real work behaviors. Agents are initialized with realistic employee profiles (role, ID, system context), enabling them to act like real staff members. To simulate insider threats, some agents are designated as adversaries—executing attacks while still performing routine tasks—making scenarios both realistic and challenging.

Organizational Operations & Threat Simulation

With the environment and agents in place, Chimera simulates daily organizational workflows. Benign agents follow task schedules, attend meetings, and update plans dynamically through interactions—each running in isolated Docker containers with realistic access-control rules. Adversarial agents blend malicious activities into their normal routines, executing attacks at contextually appropriate times while maintaining regular duties. All attack actions are pre-defined, scheduled, and automatically labeled, ensuring realistic yet well-annotated logs for downstream research.