Prompt injection protection

ACOG screens prompts for attempts to override a model’s instructions: instruction overrides, role and persona hijacks, attempts to extract the system prompt, guardrail evasion and delimiter injection.

Screening is on by default, in log-only mode. Attempts are recorded and requests go through unchanged until you choose to block.

Settings

Open Admin → Governance → Prompt security. Changing the setting requires Manage data-loss-prevention rules and connect source code (admin:security).

Setting	Effect
Screen prompts for injection attempts off	Prompts are sent with no screening, and nothing is recorded
On, Log only	Every prompt is screened; attempts are recorded and the request continues
On, Block the request	Every prompt is screened; an attempt is refused before it reaches the model

The mode can also be set per team and per API key. The most specific setting wins: key, then team, then organization.

What a blocked request gets

{
  "error": {
    "type": "prompt_injection_blocked",
    "message": "…",
    "category": "instruction_override"
  }
}

The status is 400. Blocks are delivered as the governance.guardrail_blocked webhook event.

Reviewing attempts

The Injection attempts log on the same page lists every attempt caught, with its category and who sent it. It is visible to holders of admin:security.

Where it applies

Prompt screening runs on the compatible endpoints and the native routes. It does not run on embeddings requests.