Agentic Chaos Monkey
Agentic Chaos Monkey is an autonomous fault-injection and compliance-auditing framework purpose-built for AI agent infrastructure.
Core trick
A proxy intercepts tool-call JSON and mutates it with a saboteur agent, while inline guardrails block bad payloads before the agent sees them.
Description
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.
venkatacrc/chaos-monkey-distributed-agents★ 00 forksHTMLCreated May 23, 2026 · last push May 24, 2026
README excerpt
# Chaos Engineering Framework for Agentic Infrastructure 🐒⚡ A chaos engineering and resilience testing framework purpose-built for autonomous AI systems. The platform intercepts, mutates, and injects runtime anomalies into live tool-output streams to stress-test the decision-making, safety boundaries, and fault tolerance of multi-step AI agents at scale. Built using a multi-agent orchestration architecture powered entirely by **Gemini 3.5 Flash**. --- # Overview Modern AI systems are evolving from simple chat interfaces into autonomous agents capable of executing multi-step workflows, invoking tools, and making state-changing decisions. This introduces an emerging failure class: ## Cascading Data-Stream Failures Traditional security tooling focuses on: * Static code analysis * API/network inspection * Infrastructure monitoring However, these systems often fail to detect **semantic corruption inside runtime tool payloads**. Examples include: * Converting numeric values into strings * Injecting invalid account balances * Manipulating enum states * Returning structurally valid but logically corrupted JSON When autonomous agents consume compromised payloads across multiple execution steps, the result can be: * Silent financial leakage * Unsafe automation decisions * Logic bypasses * Agent crashes * Corrupted downstream actions This framework acts as a **Chaos Monkey for AI agents**, actively attacking the information layer to uncover vulnerabilities before deploymen