Injection Guard

Prompt injection is an attack where malicious content embedded in an external source (a web page, a document, an API response) instructs the agent to take unintended actions. Reeboot’s injection guard detects and flags this content before it influences the agent.

Injection guard is enabled by default.

How It Works

When the agent calls a tool listed in external_source_tools (default: fetch_url, web_fetch), the injection guard scans the returned content for patterns that look like system-level instructions attempting to override the agent’s behaviour.

Detected content is flagged and the agent is warned before processing it. If the channel’s trust level is "end-user", the guard is applied more aggressively.

Configuration

{
  "security": {
    "injection_guard": {
      "enabled": true,
      "external_source_tools": ["fetch_url", "web_fetch"]
    }
  }
}

Adding Custom External Tools

If you add an MCP server tool that fetches external content, add it to the list:

{
  "security": {
    "injection_guard": {
      "external_source_tools": ["fetch_url", "web_fetch", "my_mcp_fetch_tool"]
    }
  }
}

Interaction with Trust Level

• trust: "owner" (default): injection guard is applied but with lighter scrutiny — the owner is expected to request fetches from known sources.
• trust: "end-user": injection guard is applied with full scrutiny — external users may direct the agent to fetch attacker-controlled content.

→ See Trust and Access Control and Permission Tiers.