Lakera Guard alternative for PII redaction in LLM apps

Lakera Guard is a screening API. You send it the messages of a conversation and it tells you whether it found a threat; for personal data it can return where each finding is, so your application blocks the call or masks the text. It does not forward the call to the model, and the documentation we read describes no step that puts the real values back into the answer.

Nonimo Gateway runs on your own servers. See plans and get in touch.

Lakera Guard is AI security in one API: prompt attacks and agent traffic, including tool responses, tool descriptions and tool-call data leakage. If your main worry is client names and numbers reaching an LLM provider, and staff still need a readable answer with the real values in it, you need a proxy that pseudonymises on the way out and restores on the way back. That is what Nonimo Gateway does.

What Lakera Guard is today

Check Point announced the agreement to acquire Lakera on 16 September 2025 and closed it on 22 October 2025. The documentation now names the product AI Guardrails, though “Lakera Guard” still appears in platform headings, which is why people keep searching for it under the old name.

The core call is POST https://api.lakera.ai/v2/guard with a messages array of role and content objects. The answer is a verdict. Your application decides what to do with it, and your application calls the model. The docs do not describe the endpoint as OpenAI-compatible.

Where Lakera Guard stops for personal data

Your code does the masking, and nothing restores it

Lakera’s documentation says it can “block the interaction or mask sensitive information”. For masking, the API returns the location and type of each finding, which “can then be used by your application to mask sensitive contents”. So:

• Your code replaces the values, using the locations Lakera returns.
• Your code chooses what to replace each value with.
• If you want the real values back in the model’s answer, your code keeps a mapping and reverses it. The documentation we read describes no restore step.

Eight built-in types, three of them US-only

The built-in PII types are full names, US mailing addresses, US phone numbers, email addresses, IP addresses, credit card numbers, IBANs and US Social Security numbers, plus custom entities and regular expressions. The documentation says non-US postal addresses and phone formats are not recognised. A Spanish NIF, a German tax ID or an Italian codice fiscale would need a custom entity that you write and maintain.

Streaming is screened, not restored

The guard API page describes splitting a streamed answer into chunks and screening each one, holding output back briefly or pulling it back if a later chunk is flagged. That is a pattern you build in your application, and it screens the stream rather than putting values back into it.

Where it runs

The hosted API has EU, US and Singapore endpoints. Self-hosting needs an Enterprise licence and runs on Kubernetes with a Helm chart or on Docker, with offline deployment supported. Check Point says it can “use NVIDIA GPUs for low-latency inference” and keeps hardware sizing off the public page.

Lakera Guard and Nonimo Gateway, side by side

	Lakera Guard (Check Point AI Guardrails)	Nonimo Gateway
What it is	Screening API that returns a verdict	OpenAI-compatible proxy in front of the API
Main purpose	AI security: prompt attacks, agents, data leakage	Personal data in prompts and answers
Who calls the model	Your application	The gateway, with your own OpenAI key
Personal data action	Block, or masking done by your code	Replaces values in the request
Real values back in the answer	Not described in the docs	Yes
Built-in PII types	8; addresses, phones and SSNs US-only	National documents and numbers of each country covered
Where it runs	EU, US or Singapore regions; self-hosted on Enterprise	On your own servers
Agent and tool screening	Yes, documented	Not its focus

What the difference looks like for an accountancy

Take a typical request from a 30-person accountancy: summarise this client email and draft a reply. The email names the client, gives a Spanish phone number, quotes a NIF and mentions an IBAN.

With Lakera Guard, the name and the IBAN fall within built-in types, the Spanish phone number falls outside the US-only detector, and the NIF needs a custom rule. Block, and the user gets no draft. Mask, and your code sends the cleaned text and gets back a reply that says “Dear [NAME]”; turning that into a letter to the actual client is code your team writes.

With Nonimo Gateway, the request leaves with stand-ins for the client’s name, the NIF and the other personal data it detects, the model drafts around them, and the gateway puts the real values back in the answer before it reaches your application. Reversible replacement is pseudonymisation, not anonymisation; pseudonymisation under UK GDPR explains what that means.

How Nonimo Gateway does it

Nonimo Gateway runs on your own servers. It exposes an OpenAI-compatible API, so the change in your code is the base URL:

from openai import OpenAI

client = OpenAI(
    base_url="https://gateway.your-company.internal/v1",  # was https://api.openai.com/v1
    api_key="sk-...your own OpenAI key...",
)

It replaces personal data in what goes out and restores the real values in the response. It recognises the national documents and numbers of each country it serves, such as the Spanish DNI, NIE and NIF, Spain’s CUPS energy supply number, the German Steuer-ID, the French NIR, the Italian codice fiscale, the Brazilian CPF and the Mexican RFC and CURP. There is a plugin for LiteLLM if you already route through it, and you keep your own OpenAI key. Plans and prices are on the Nonimo Gateway page.

It helps with pseudonymisation; it does not make a provider’s processing lawful on its own. The legal side is in is putting client data in ChatGPT a data breach, and the practical list of what to take out in what to redact before you paste a document into an AI.

If your problem is client data in LLM calls rather than prompt attacks, Nonimo Gateway is built for it. See plans and get in touch. For the rest of the market, see the overview of PII redaction software, or the comparisons with Cloudflare AI Gateway and LLM Guard.

Sources

• Check Point, press release on acquiring Lakera, 16 September 2025. The agreement to acquire.
• Check Point, third-quarter 2025 results, 28 October 2025. The acquisition closed on 22 October 2025.
• Lakera, homepage. Check Point branding, no published prices.
• Lakera, API documentation. The guard endpoint, message format, EU, US and Singapore regions, tool and agent screening.
• Lakera, guard API reference. Finding locations for masking by your application, chunked streaming pattern.
• Lakera, data leakage prevention. Block or mask, the eight PII types.
• Lakera, self-hosting. Enterprise licence, Helm and Docker, offline deployment, NVIDIA GPUs for low latency.
• Lakera, platform pricing page. No prices shown when checked.

All checked 20 October 2026.

---

This page is for information. Product details of other vendors were checked on their official pages on 20 October 2026 and can change.