AgentChaos

Agentic Chaos Monkey is an autonomous fault-injection and compliance-auditing framework purpose-built for AI agent infrastructure. Utilizing a multi-agent orchestration architecture powered by Gemini 3.5 Flash, the framework: Intercepts raw JSON tool payloads at runtime in a proxy layer. Mutates payloads dynamically using a malicious Saboteur Agent to inject realistic type coercion, negative boundaries, or custom semantic exploits. Defends runtime state via programmatic Shield Guardrails that act as an inline firewall. Programmatically evaluates agent resilience using a Compliance Evaluator to analyze logs and stream back a formal security scorecard. All of this is visualized on a gorgeous, responsive, real-time node dashboard served directly on Google Cloud Run, demonstrating when agents fail, when they safely abort, and when they are shielded.