Agenstral is security tooling, but it is not a sandbox. A denied call is blocked before execution when routed through an Agenstral proxy. A call that bypasses Agenstral is outside its control.

This project is pre-1.0. Security fixes target the default branch until formal releases begin.

Please open a private security advisory if the repository host supports it. If not, open a minimal public issue that avoids exploit details and request a private contact path.

Agenstral is designed to:

• Discover risky agent and MCP configuration.
• Enforce local policy before proxied tool calls execute.
• Detect secret-looking values before logging or reporting.
• Produce tamper-evident audit records for local verification.

Agenstral is not designed to:

• Contain malicious processes after execution.
• Replace OS sandboxing, container isolation, or network egress controls.
• Prove that no unobserved agent action happened outside the proxy path.