Hey everyone! 👋
I built a Zero-Trust AI Agent Security Engine designed to validate AI system workflows, prevent prompt injection, and guard automated agents. It has already passed over 4,300+ test scenarios.
📹 Watch the video demo on X:
https://x.com/Eiyaso/status/2093772627903197513
Would love to hear your feedback from the developer community
Top comments (1)
I’ve been building something I believe will become increasingly important as AI agents become more autonomous.
AI agents are no longer limited to generating text. They can call APIs, access databases, execute tools, interact with MCP servers, modify files, browse the web, and perform multi-step workflows.
That creates a different security problem:
How do we verify that an AI agent is behaving safely throughout its entire execution trajectory — not just whether the initial prompt is safe?
I’ve been working on a prototype Zero-Trust AI Agent Behavior Verification Engine designed around this problem.
The system focuses on:
• Tool-call authorization and policy enforcement
• Agent trajectory / multi-step behavior verification
• Prompt-injection and indirect-injection detection
• MCP gateway security and schema validation
• Credential/secret exposure detection
• Privilege-escalation protection
• Unauthorized tool and resource access
• Memory / vector-store protection
• Automated anomaly detection
• Code-level patch generation and validation
• Sandboxed testing of proposed fixes
• Security monitoring and observability
• Automated regression testing
The architecture is modular and built with Python, FastAPI, Streamlit and MCP-related security concepts.
I’m currently looking for security engineers, AI infrastructure teams, founders, researchers and companies working on agentic AI who are interested in reviewing the architecture and challenging the system with real-world adversarial scenarios.
I’m especially interested in feedback on:
“What happens when an otherwise legitimate AI agent gradually becomes dangerous through a sequence of individually valid actions?”
I’m building this publicly, testing it continuously, and looking for people who can break it rather than simply tell me it looks good.
Technical feedback, security critiques and collaboration opportunities are welcome