Prompt Injection Blocking

Prompt injection blocking scans incoming payloads for patterns that attempt to override or manipulate the agent’s instructions.

Configuration

policies:
  safety:
    block_prompt_injection: true

Field	Type	Default	Description
block_prompt_injection	bool	true	When true, input is scanned against built-in injection patterns before invocation

This is enabled by default whenever a safety block is present.

Built-in Patterns

The RuntimePolicyEngine in dockrion_runtime/policies.py maintains a compiled list of regex patterns (_injection_patterns) that match common prompt injection techniques:

• “ignore previous instructions” variants
• “you are now” role reassignment attempts
• “system:” or ”### system” prompt override patterns
• Attempts to extract system prompts
• Common jailbreak patterns

These patterns are compiled at startup for performance. The exact list is defined in the RuntimePolicyEngine.__init__() method.

How It Works

1. Before the adapter’s invoke() is called, validate_input() runs on the serialized input
2. Each string value in the input payload is checked against all injection patterns
3. If a match is found:
   • With halt_on_violation: false (default) → the request is rejected with a PolicyViolationError
   • With halt_on_violation: true → same behavior (injection blocking always rejects; halt_on_violation primarily affects output policies)

The check is performed by RuntimePolicyEngine.validate_input(), which serializes the payload to JSON and runs regex matching against the entire string.

Limitations

• Pattern-based detection is not foolproof — sophisticated injection attacks may bypass regex patterns
• No ML-based injection detection is included (future consideration)
• Only string values in the payload are checked; binary or encoded content is not analyzed

Disabling

If your agent handles untrusted input by design (e.g., a red-teaming tool):

policies:
  safety:
    block_prompt_injection: false

Source: SafetyPolicy.block_prompt_injection in packages/schema/dockrion_schema/dockfile_v1.py; RuntimePolicyEngine._injection_patterns in packages/runtime/dockrion_runtime/policies.py

---

Up: Policies Overview | Next: Output Controls →