# Security and agent safety

Humans stay in control of high-risk actions. Sensitive or transactional actions require explicit human approval, authentication, and scoped permissions.

## Principles

- Agent-facing tools are read-only by default.
- Sensitive or transactional actions require explicit human approval, authentication, and scoped permissions.
- When an agent attempts an action, the interface shows what data the agent accessed and what action it wants to take.
- Payment, email, and record modification are never completed without explicit authenticated approval.
- WebMCP is experimental and is positioned as an early-adopter enhancement, not a universal standard or guaranteed agent channel.
- AI Agentship does not claim security or compliance certifications unless already true.

Source: https://www.aiagentship.com/security.md
Updated: 2026-02-14