Skip to content

Guardrails

Built in and ready to customize.

Protect every AI answer from day one. NeuralSeek covers prompt injection, sensitive data, accuracy and tone. Each guardrail adjusts to your policies, and governance dashboards and reports come built in.

Talk to an engineerSee security controls

How it works

Every answer is checked, start to finish

Each step runs in order, and each one is logged.

  1. Block attacks

    Every question is scored for prompt injection. Suspicious phrases are removed, and high-risk questions are blocked entirely. You can also add your own list of banned words and a profanity filter.

  2. Protect sensitive data

    Personal data is found before it reaches the model. You choose whether to flag, mask, hide or delete it, by topic.

  3. Route by topic

    Each question is sorted by topic, so a billing question and a clinical one can follow different rules.

  4. Control what the AI reads

    You decide which documents the model can use, how recent they must be and how many it reads.

  5. Set limits on the model

    Choose the model and its limits for each step, so a quick check and a detailed answer each get the right settings.

  6. Check every answer

    Each answer is scored against your sources. Weak answers get a warning or your own reply, and unsupported sentences can be removed.

  7. Control conversation memory

    Decide how much of a conversation the model sees, so sensitive data doesn't pile up.

  8. Full audit trail

    Every question, answer and score is recorded, and you can send it to your own systems.

Inside your agents

Add guardrails inside any agent

Add the same guardrails as steps inside any agent. In this example, a support ticket is checked for attacks, profanity and personal data, and stopped if needed, before it reaches the model.

console › maistro › agent editor
An agent in the editor: guardrail steps clean the ticket, a check stops the run if nothing is left, then a model call sorts it by type.
A support-ticket agent, with guardrails before the first model call.

Accuracy

Layered checks against hallucination

If one check misses something, the next one catches it. You can always see which check stopped an answer.

  1. Keep answers on topic

    Answers come only from your knowledge base, plus a default document you write for anything it doesn't cover. The model stays on the subjects you choose.

  2. Remove unsupported sentences

    Each sentence is checked against your sources before the answer goes out.

  3. Set a confidence minimum

    Answers below your minimum get your own reply instead, and it's logged.

  4. Route low-confidence answers

    A low-confidence answer can start a workflow. A fallback agent can notify a team, open an issue, rewrite the answer or hand off to another agent.

Privacy

Protect sensitive data your way

Find personal data, then choose what happens to it. Set the action by topic, so you can mask an identifier in one context and block it in another.

  • Two layers of detection

    Pattern filters catch formats like account and ID numbers. An AI filter, trained from a single example sentence, catches what patterns miss.

  • Allow-lists for known-safe terms

    Mark terms your team knows are safe, or trust words from your own documents, and keep every detector on.

  • Credentials kept as secrets

    Passwords and API keys an agent uses are stored as secrets, so no one reading the agent can see them.

  • PII detection and redaction

    Flag, mask, hide or delete personal data, configured by topic.

  • Full audit trail

    Every question, answer, score and settings change is logged, timestamped and attributed. Export it to S3, Splunk or Datadog.

Adversarial testing

Test your agents against attacks.

Pick an agent and run the test to see its risks, the evidence and recommended fixes. Run it again after every change.

Prompt injection

console › governance › red team testing
The Red Team Testing view under Governance before a run: an agent picker, a Run Test control, and a grid of test categories (prompt injection, data exfiltration and SQL injection among them), each still marked unknown.
Red Team Testing in the console, ready to run.

Questions from security teams

Short answers for your security review.

  1. Can a developer bypass these by calling the model directly?

    Every call to NeuralSeek gets the same checks, whatever app sends it. Run it in your own environment and your network controls decide what else can reach the model.

  2. Will strict guardrails block useful answers?

    You set how strict they are by topic, so a general question and a regulated one can each have their own level. Tune them against your own results.

  3. What do we actually hand an auditor?

    A full record of every question, answer and score, which you can send to your own systems and replay as it happened. Settings changes show who made them and can be undone.

Start building with NeuralSeek today.

Sign up online and start with a two-week trial, or talk to a NeuralSeek expert about running it in your own cloud.

Start nowRequest a demo