Pryxor

Pryxor: A Runtime Security Layer for AI Agents

Pryxor provides a robust runtime authorization mechanism for AI agent actions, ensuring that even with valid credentials, AI agents can only execute the actions that are deemed safe. This self-hosted runtime gateway critically evaluates each proposed action before it interacts with the underlying systems.

Key Features

• Action Evaluation: Pryxor assesses an agent's proposed action against a set of defined policies before any interactions with external systems occur.
• Robust Decision-Making: It can approve safe actions, block forbidden actions, or pause risky actions for human oversight.
• Credential Isolation: Actions are executed without exposing the agent to any real credentials, enhancing the security posture.
• Action Governance: Each action's approval process is recorded, creating an auditable lifecycle for security compliance.

The Problem: Authorized but Dangerous

Authorization typically focuses on whether an identity can make an API call. However, Pryxor addresses a more critical question: should this specific action be allowed at this moment? An agent that can read records from a database might inadvertently access or export sensitive data beyond its intended task. Pryxor aims to provide an explicit enforcement mechanism for such scenarios, significantly reducing the risk associated with unintended actions.

How Pryxor Works

When an AI agent proposes an action, Pryxor:

1. Authenticates the agent.
2. Validates the action arguments.
3. Evaluates the action against its policies.
4. Returns one of three outcomes: APPROVED, HOLD, or BLOCKED.

Policy Enforcement Example

Pryxor supports explicit policies governing actions. For instance:

{
  "type": "declarative",
  "rules": [
    {
      "name": "allow-internal-emails",
      "when": {
        "tool": "send_email",
        "to_domain_in": ["company.local"]
      },
      "then": {
        "approve": true
      }
    },
    {
      "name": "hold-external-emails",
      "when": {
        "tool": "send_email"
      },
      "then": {
        "hold": "EXTERNAL_EMAIL",
        "message": "External recipients require human approval."
      }
    }
  ]
}

Policies are evaluated prior to action execution, ensuring only compliant actions are allowed.

Current Limitations and Future Directions

Pryxor is in the early stages of development, suitable for local evaluations and security experiments. Continuous enhancements are planned to ensure better policy semantics, reliable action execution, and enhanced operational capabilities.

Contribution Opportunities

Pryxor welcomes contributions in the form of test cases, policy suggestions, executor integrations, and security reviews. Contributions are essential in refining the system's capability and effectiveness.

Pryxor aims to provide a crucial layer of security for AI agents, ensuring that actions are not just authorized but are appropriate and safe, helping organizations effectively manage their AI-driven workflows.