Runtime guardrails API that screens prompts, model output and agent tool calls, with EU, US and Singapore processing regions and a self-hosted option.
Platform and security engineers who want one screening call in front of an LLM or agent and need the processing region pinned, since the Guard endpoint is published per region. It suits cases where a single policy must cover conversational prompts, retrieved content, tool responses sent with the tool role and MCP tool descriptions, because all of them go through the same POST to /v2/guard. It also fits teams already running managed agent platforms, as the docs name connectors for Amazon Bedrock Agents and AgentCore, Google Cloud's agent registry, Microsoft Copilot Studio, Salesforce Agentforce, n8n and Relevance AI.
Pricing is quote-only: the vendor's pricing link resolves to a gated platform page and no tiers, request allowances or seat prices are published, so any budget needs a sales conversation first. Self-hosting is not something you can buy online either, because the docs require a valid AI Guardrails Enterprise licence plus container registry credentials issued by Check Point, and the self-hosted latency benchmarks and sizing guidance sit behind a separate customer portal rather than the public docs. Coverage is narrower than the marketing suggests: the Prompt Defense page says the detector is optimised for conversational AI and discusses text only, and the guardrails page states that the detectors can produce both false positives and false negatives.
Pros
Cons
The marketing site states it is 'Delivering sub-50 ms runtime latency' and separately promises minimal latency even for very large prompts and context windows. Both are vendor claims with no published methodology or percentile breakdown. The integration docs add a practical note: screening inputs sequentially before the model call adds latency, so they recommend running input screening in parallel with inference and using a single screening request per interaction.
Processing follows the endpoint you call: eu.api.lakera.ai maps to eu-south-2, us.api.lakera.ai to us-east-1, and ap-southeast-1.api.lakera.ai to Singapore, while api.lakera.ai routes to the nearest. Requests are not auto-routed, so you choose. Log storage defaults to the EU and can be set to the US, but it is fixed at organisation creation and cannot be changed; two storage regions means two organisations.
The site claims risks are reduced by three to four orders of magnitude, cites a 0.01% production false positive rate attributed to a customer, and says it observes over 100 languages in real time. None of these carries published methodology or independent verification. The guardrails documentation is more cautious, stating plainly that detectors can produce both false positives and false negatives, and that threshold levels L1 to L4 trade one against the other.
Yes. The docs cover Kubernetes via a Helm chart, Docker or any OCI-compatible runtime, and air-gapped installation where containers are exported and loaded offline, with optional NVIDIA GPU acceleration and Kubernetes health probes. It is not self-serve: you need a valid AI Guardrails Enterprise licence plus container registry credentials issued by Check Point, and the latency benchmarks and sizing guidance live in a separate customer portal.
Facts sourced from the vendor’s own site, last verified 2026-10-10.
The Weekly Source
One email a week. New listings, repriced tools, and the ones we would skip - including this one if its pricing changes.
