What it is
Lakera Guard is a screening API for LLM applications, now owned by Check Point and documented as Check Point AI Guardrails. Check Point bought Lakera in 2025, and the docs, terms and privacy paperwork now carry its name. Your application sends each user message, retrieved document, tool response or model output to one endpoint, POST /v2/guard, and gets back a flagged result it can act on.
The managed guardrails cover prompt attacks (direct and indirect injection and jailbreaks), content moderation, data-leak prevention for personal data and system-prompt leakage, malicious-link detection and an agent-behaviour defence with a tool allow and deny list. Custom guardrails written in plain language are in beta. Each detector runs at one of four strictness levels, L1 lenient to L4 paranoid, and L4 is the default.
The free Community plan gives 10,000 requests a month on the hosted service, with prompts up to 8,000 tokens. Enterprise, priced by quote, adds self-hosting on Kubernetes or Docker (including air-gapped installs), SSO, role-based access, SIEM integration and a choice of EU or US data residency.
Our review
Lakera Guard does one job properly: it puts a maintained classifier for prompt injection and data leakage between your users and your model, without you training or hosting anything. The API takes messages in OpenAI's chat format, including tool calls and tool definitions, and screens only the last interaction while treating earlier turns as context. That design suits agents, where the dangerous text usually arrives in a retrieved page or a tool response rather than from the user.
The defaults deserve attention before launch. Every detector ships at L4, the paranoid setting, and the docs say plainly that higher levels buy fewer misses with more false positives. The project-less default policy is described as "intentionally strict". A team that goes live without tuning will block legitimate traffic, so the sensible path is Detect mode, which logs without blocking, on a sample of real traffic first. Allow and deny lists can patch individual errors, but Check Point's own docs warn they can open security holes.
Data handling is clearer than most rivals because Check Point publishes a privacy data sheet. On the hosted service it processes prompts, system prompts, tool information, model outputs and prompt embeddings, and keeps them for the life of the subscription plus three months. Customers can delete prompts and restrict retention of pseudonymised prompts. Logs are stored in the EU unless you pick another region when you create the organisation, and that choice cannot be changed later. Self-hosted deployments keep data inside your environment.
The commercial picture is the weak point. Lakera publishes the free tier and nothing else: no per-request rate, no rate limits, no latency figure beyond its own sub-50 ms marketing claim. Each screened item counts as a request, so an agent that checks the user prompt, two retrieved documents and the final answer spends four requests per turn and exhausts the free 10,000 at 2,500 turns. NeMo Guardrails or a classifier such as Llama Guard costs nothing in licences but leaves you owning the models, the GPUs and the false-positive tuning that Lakera sells.
Where it earns its place
A backend engineer adding a support chatbot with RAG over help-centre articles
Stop injected instructions in retrieved pages from steering the bot
Send the retrieved chunks and the user turn to /v2/guard before the model call, and the answer again before display. Start in Detect mode to measure false positives at L4, then loosen detectors that misfire. Free up to 10,000 requests a month, which is roughly 3,300 conversations if each turn needs three screens.
A platform team at a bank running internal agents with tool access
Keep screening on-premises and restrict which tools an agent may call
An Enterprise licence unlocks the Helm chart or Docker image, which can run air-gapped with NVIDIA GPUs; data then stays in the bank's environment. The tool allow and deny list blocks calls outside an agent's mandate. Pricing and hardware sizing come only from Check Point sales.
A compliance lead at a 200-person EU software company
Approve a hosted guardrail vendor under GDPR
Use the EU processing endpoint and keep log storage in the EU, set at organisation creation and not changeable. The privacy data sheet lists retention of the subscription term plus three months and names AWS hosting; Check Point's SOC 2 and GDPR claims sit on the Enterprise tier.
Before you commit
- Setup
- Create a project, attach a policy and call POST https://api.lakera.ai/v2/guard with a bearer key; the quickstart is minutes, but tuning strictness on real traffic is the real work.
- Integrations
- REST API with an OpenAPI spec; regional endpoints for the EU, US and Singapore; platform API for policies, projects, logs and analytics; Helm chart and Docker images for self-hosting.
- Data and privacy
- Hosted service keeps prompts, outputs and embeddings for the subscription plus three months on AWS in the EU, US or Singapore; customers can delete prompts and limit retention. Self-hosted data stays with the customer. SOC 2 and GDPR are listed for Enterprise.
- Cost at scale
- Not published beyond the free 10,000 requests. An agent screening four items per turn at 100,000 turns a month needs 400,000 requests, which means an Enterprise quote.
- Limits
- Community prompts are capped at 8,000 tokens; only text content is screened; rate limits are not published; streaming output should be screened in chunks of at least 10 tokens.
What's good
- One endpoint screens prompts, retrieved documents, tool output and replies
- Accepts OpenAI chat format, including tool calls and definitions
- Four documented strictness levels per detector, plus a Detect-only mode
- Published retention: subscription term plus three months
- Air-gapped self-hosting on Enterprise
What's not
- Defaults to L4, the strictest level, so expect false positives untuned
- Only the free tier is priced; everything else is a sales call
- Rate limits and latency figures not published
- Log storage region is fixed at organisation creation
- Self-hosting, SSO and SIEM are Enterprise-only
Best for
Teams that want a managed screen for prompt injection and data leaks in front of chatbots and agents, without operating their own classifiers.
Alternatives
Lakera Guard (Check Point AI Guardrails): common questions
- Is Lakera Guard (Check Point AI Guardrails) worth it?
- Lakera Guard is a well-documented injection screen you can wire in with one API call, but past 10,000 requests a month you are negotiating with Check Point's enterprise sales team. We rate it 4.0 out of 5.
- How much does Lakera Guard (Check Point AI Guardrails) cost?
- Community free: 10,000 requests/month, 8k-token prompts, SaaS only. Enterprise by quote: flexible volume, SaaS or self-hosted, SSO, RBAC, SIEM integration. No published price beyond the free tier.
- What are the downsides of Lakera Guard (Check Point AI Guardrails)?
- Defaults to L4, the strictest level, so expect false positives untuned; only the free tier is priced; everything else is a sales call; rate limits and latency figures not published; log storage region is fixed at organisation creation; self-hosting, SSO and SIEM are Enterprise-only.
- Who is Lakera Guard (Check Point AI Guardrails) best for?
- Teams that want a managed screen for prompt injection and data leaks in front of chatbots and agents, without operating their own classifiers.
- How much does Lakera Guard (Check Point AI Guardrails) cost for a team or in production?
- Not published beyond the free 10,000 requests. An agent screening four items per turn at 100,000 turns a month needs 400,000 requests, which means an Enterprise quote.
- Is Lakera Guard (Check Point AI Guardrails) safe for company data?
- Hosted service keeps prompts, outputs and embeddings for the subscription plus three months on AWS in the EU, US or Singapore; customers can delete prompts and limit retention. Self-hosted data stays with the customer. SOC 2 and GDPR are listed for Enterprise.
- What are the best Lakera Guard (Check Point AI Guardrails) alternatives?
- The closest alternatives we have reviewed are Amazon Bedrock and Google Agent Platform (Vertex AI).
Answers checked Oct 23, 2026.
Make Lakera Guard (Check Point AI Guardrails)? Show this rating on your site
