Lakera Guard

Check Point Software Technologies (NASDAQ: CHKP) announced an agreement to acquire Lakera on 16 September 2025; the deal was reported at roughly $300 million by trade press, though Check Point has not disclosed terms itself. As of today, lakera.ai/about redirects (HTTP 301) straight to checkpoint.com/about-us/, and the dedicated lakera.ai/lakera-guard product page redirects to /ai-agent-security. Lakera's own documentation site now serves the Check Point logo, and the site footer reads "Lakera Inc... ©1994-2026 Check Point Software Technologies Ltd." Every public repository under the lakeraai GitHub organization (pint-benchmark, chrome-extension, onnx_clip, canica, intent-augmentation, dsec-gandalf, guard-demo-client) is now archived; active development on at least the PINT prompt-injection benchmark continues under a new CheckPointSW/pint-benchmark fork, with commits into April 2026. A buyer should evaluate this as Check Point's roadmap and support commitment, not Lakera's as a standalone company — the founding team's continuity and independence claims that mattered pre-acquisition no longer apply.

Guard's job is inline detection on the request/response path of an LLM application or agent: it screens prompts and model outputs for prompt injection, jailbreak attempts, and sensitive-data exposure, and (per the newer "AI Agent Security" framing) extends into agent-specific risks like unsafe tool calls and MCP-connected system abuse. It ships as an API you call inline, not a passive log-and-alert tool. Lakera publishes an open PINT (Prompt Injection Test) benchmark and dataset on GitHub, and its own last-published run (2 May 2025) put Lakera Guard at 95.22% versus 89.24% for AWS Bedrock Guardrails, 89.12% for Azure AI Prompt Shield, and 70.07% for Google Model Armor. This is Lakera's own benchmark, over methodology and a dataset it also authored and maintains, so treat the ranking as a self-reported result rather than independent verification — though the dataset and evaluation code are public and open to community scrutiny, which is more transparency than most vendors offer on detection accuracy. The score is also 16 months old at the time of writing; there's no newer published PINT run to check whether Guard or competitors have shifted since.

As of September 2026, Lakera Guard/Check Point AI Agent Security has two published tiers. Community is $0/month: 10,000 requests/month, an 8,000-token maximum prompt size, SaaS-only hosting, community support, API access, dashboards, reports, encryption in transit and at rest, SOC 2 and GDPR compliance, and EU data residency. Enterprise has no published price — "flexible" request volume, configurable prompt size, SaaS or self-hosted deployment, enterprise support, SSO, role-based access control, SIEM integration, version pinning (self-hosted only), and EU/US data residency, all behind a sales conversation. There's no paid self-serve middle tier: teams that outgrow the free 10k requests/month go straight to a custom quote, which is worth knowing before you build an integration assuming a metered-usage plan will exist.

Both pricing tiers claim SOC 2 and GDPR compliance, and Check Point maintains a public trust center (checkpoint.com/trust-point/) listing named policies — information security, physical security, business continuity/disaster recovery, vendor security assessment, vulnerability disclosure, and data-subject-rights procedures — that now cover Lakera as part of the broader Check Point platform. A group-wide status page exists at status.checkpoint.com. There is no Lakera-specific security or compliance page anymore; what used to be lakera.ai/security is now a general Check Point AI-security marketing page, not a trust/compliance document.

Gandalf, Lakera's public prompt-injection game (play.lakera.ai), is still live and still linked from the main site, as is a ~1,000-member Slack community ("Momentum"). These predate the acquisition and appear to have continued under Check Point, which is a reasonable signal that the acquirer sees ongoing value in Lakera's research/community brand rather than shutting it down for parts.