Skip to content
Book a demo
Menu
Book a demo

AI Security · AI Firewall

Control What Goes Into AI — and What Comes Back

AI interactions carry risk in both directions. GuardStance AI Firewall inspects prompts going out and model responses coming back in real time—masking sensitive data, catching secrets, and enforcing security policies across your AI workflows.

See AI Firewall in Action

A support ticket goes to an LLM.

Data
Ticket text with a customer’s IBAN
User
Support agent in Zendesk
Destination
LLM through the AI gateway
Policy
Financial data protection — mask before sending
Decision
Mask IBAN masked before transmission; original value restored for the authorized user in the response
Audit
Event recorded

When a Prompt Contains More Than a Question

When an employee pastes a customer record, financial figure, or source code into a prompt, GuardStance detects the sensitive content before it ever leaves the browser, endpoint, or gateway.

Depending on your policy, GuardStance warns the user, masks the sensitive tokens inline so the prompt can still proceed safely, or blocks the request entirely.

Where does it run?

  1. Browser extension

    Inspects prompts, file uploads, and responses across browser-based AI services like ChatGPT, Claude, and Gemini.

  2. Endpoint agent

    Monitors and controls desktop AI applications, local coding assistants, and clipboard pastes.

  3. AI gateway

    Inspects AI requests and responses routed through the gateway from integrated business applications such as Salesforce, Teams or Zendesk to models such as OpenAI, Gemini or Mistral.

What does an AI firewall check?

On the way in: prompts and files

  • Personal data masking

    Names, IDs and account numbers are swapped for labels. The prompt still goes out.

  • Secrets and tokens

    API keys, passwords and tokens are caught before they reach a model.

  • Source code

    Code pasted into a prompt is recognized and handled by policy.

  • Prompt injection and jailbreak

    Detect prompt injection and jailbreak patterns before they reach the model and enforce your security policy.

On the way out: answers

  • Masked values restored

    The user sees the real values again in the answer. The model never did.

  • Sensitive data in answers

    Answers that carry protected data are caught on the way back.

  • Sensitive content and policy restrictions

    Prevent models from returning restricted internal data or policy-violating outputs to the user.

Questions about the AI Firewall

Do people have to stop using AI tools?

No. Instead of blanket bans, you can tailor responses by data sensitivity and user context—allowing employees to continue working safely after inline data masking or a justification prompt, while blocking high-risk leaks.

Which AI services does it cover?

GuardStance covers browser-based LLMs (including ChatGPT, Claude, Gemini, Perplexity, and DeepSeek) via the browser extension, desktop AI apps and coding assistants (Copilot, Cursor, Claude Code, Codex) via the endpoint agent, and enterprise API traffic (OpenAI, Anthropic, Mistral, Gemini) via the AI Gateway. We share the full per-service control matrix before your POC.

What does an AI Firewall inspect?

GuardStance AI Firewall inspects both outbound prompts (including file attachments) and inbound model responses for PII, financial data, API keys, credentials, source code, and prompt injection patterns.

Related pages

See it on your own data.

Book a demo