Documentation / Security and Privacy

Security and Privacy

Updated October 4, 2026

Where Your Data Goes

Agent Builder runs inside your WordPress installation. Conversation history, settings, the activity log and backups are stored in your own WordPress database and wp-content folder. Data leaves your server in these cases:

  • AI provider. Prompts, and the content an agent reads to answer them, go to the AI provider you use. That is either our hosted Agentic AI service or a provider you connect with your own API key (see below).
  • Agentic AI (hosted). If you use it, we receive your site URL, licence key and prompts, and process them via Google Vertex AI / Gemini (and xAI for some voice features). Hosted chat stores no prompt text, only model, token counts, credits, status and timestamps. For image and video generation we keep the first 2,000 characters of each prompt and delete them automatically after 30 days.
  • Telemetry (self-hosted edition only, opt-in). Off by default. If you allow it, the plugin sends your site URL, licence key, plugin version and feature-usage events. The WordPress.org edition sends no telemetry.
  • Optional services you switch on. These are listed under “External Services” in the plugin readme.
Message Scanning toggle for prompt injection and malicious content detection
Message Scanning — enable security filter to scan for prompt injection and malicious content

What Goes to the AI Provider?

When you chat with an assistant, the conversation (your messages and the assistant’s responses) is sent to the AI provider you use, either our hosted Agentic AI service or one you connect yourself (OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral and others, or a local Ollama model), for processing. This includes:

  • Your message text
  • The assistant’s system instructions
  • Context the assistant gathers (e.g., post content it reads to answer your question)

Each provider has its own privacy policy:

API Key Security

Your AI provider API key is:

  • Stored in your WordPress database (encrypted using WordPress salt keys)
  • Never displayed in full in the admin interface (masked with asterisks)
  • Sent only to the provider it belongs to, never to Agentic
  • Never included in audit logs

Plugin Security Features

  • Risk-rated tools — Every tool has a risk level. In Supervised mode (the default), medium-risk actions ask for confirmation, high-risk actions wait in the Approvals queue, and extreme-risk tools are blocked
  • Approval workflows — Plain-English approval rules plus a review queue for higher-risk changes
  • Tamper-evident activity log — Agent tool calls are recorded in a hash-chained log, so an edited or deleted entry is detectable
  • Automatic backups — Before an agent changes a tracked file or database table, a snapshot is saved (last 3 per table)
  • Emergency Stop — One control that disables all agents, cancels pending jobs and disconnects AI providers until you restore them
  • WordPress capabilities — Assistants respect the WordPress permission system
  • Rate limiting — Prevents runaway API usage

Reporting a Security Issue

If you discover a security vulnerability, please report it through our Patchstack vulnerability disclosure programme or email [email protected]. See our security.txt. Please do not open a public GitHub issue. We respond within 3 business days.

Browse all documentation →