AgentGuard is a framework that autonomously identifies, validates, and constrains dangerous workflows in AI systems that use external tools, enhancing safety without human intervention.

• Automatically discovers potential attack vectors through adversarial prompting
• Validates unsafe workflows to confirm real security threats
• Generates safety constraints to restrict agent behavior at deployment time
• Provides a baseline safety guarantee for tool-using AI systems

This research is crucial for securing AI agents that can execute real-world actions through tools, preventing potentially harmful consequences from compromised systems or adversarial inputs.

AgentGuard: Repurposing Agentic Orchestrator for Safety Evaluation of Tool Orchestration