LIVE POLICY ENGINE · SAME RULES AS THE GATEWAY

Try the Shield

Inspect how a deterministic runtime policy handles prompt injection, exposed credentials and unsafe agent actions. Every example here is evaluated by the same server-side engine the hosted gateway uses, and nothing you send is stored.

LIVE POLICY ENGINE · SERVER-SIDE

deterministic

REQUEST PAYLOAD

ENFORCEMENT DECISION

—

RISK SCORE

0/100

CategoryRun an inspection to see a decision.

Matched rules—

Forwarded—

Action—

SANITIZED OUTPUT

No payload supplied.

This decision is produced by GuardBotAI’s deterministic policy engine running on our servers — the same evaluator the hosted gateway uses. Your payload is evaluated in memory and never stored.

FREE TOOL · RUNS IN YOUR BROWSER

Build your own policy, then test it here

Answer in plain language. You get a ready-to-paste GuardBot policy, an honest explanation of what each rule stops and what it costs you, and a button that sends it straight to the console above so you can watch it refuse a real attack.

What must your AI never discuss or reveal?

One phrase per line. Project names, internal terms, client names.

Which tools or functions is the agent allowed to call?

Comma or line separated. Leave empty to allow any tool (not recommended).

Largest request you want to accept (characters)

Treat suspicious requests as blocks (strict mode)Safer, with more false positives at the start.

Do you want to block, or watch first?

Which containment layer should it start on?

Your GuardBot policy

{
  "maxRequestChars": 8000,
  "strict": true
}

Enforcement mode (enforce) and containment layer (1) are set on the guardbot itself in the control center, not inside the rule set.

What each rule actually does

Built-in threat rules (always on)

Stops: Prompt injection, jailbreak phrasing, instruction override attempts, and exposed credentials such as API keys and tokens.

Trade-off: None to configure. They run on every request before anything is forwarded.

No tool allow-list

Stops: Nothing. Your agent may call any tool it names.

Trade-off: This is the single largest gap in most policies. Naming the tools your agent is allowed to use is the highest-value rule you can add.

Request size ceiling (8,000 characters)

Stops: Oversized requests, which are how long documents smuggle hidden instructions and how runaway loops burn spend.

Trade-off: Legitimate long documents will be refused. Raise the number if your workflow genuinely needs it.

Strict mode on

Stops: Anything the engine merely finds suspicious is treated as a block rather than a warning.

Trade-off: Expect some false positives early. Review the blocked events in the control center and loosen deliberately.

Enforcement mode: Enforce

Stops: Matched requests are refused with an error before a model or tool is reached.

Trade-off: Your application must handle a refusal response gracefully.

Containment layer 1

Stops: Everything. Nothing is forwarded at all — requests are judged and recorded only.

Trade-off: No real answers come back. This is the right place to start and requires no paid plan.

Create a GuardBot with this policy See a policy like this in action