Lakera

What is Lakera?

Lakera is an AI security company whose main product, Lakera Guard, sits between users and a large language model and blocks attacks before they reach it. Guard checks every prompt and response for prompt injection, jailbreaks, sensitive data leaks, harmful content and malicious links, then returns a verdict your application acts on. A second product, Lakera Red, tests AI systems with adversarial attacks before launch. Lakera also runs Gandalf, a public hacking game that it says has produced over 80 million adversarial patterns for training its detectors.

Lakera was founded in Zurich by David Haber, Dr Mateo Rojas-Carulla and Dr Matthias Kraft, with a second office in San Francisco. Check Point Software Technologies announced a deal to buy it on 16 September 2025 and said the acquisition was complete in November 2025. Zurich became Check Point’s Global Centre of Excellence for AI Security. The lakera.ai site still operates in October 2026, but the documentation now uses Check Point names such as AI Guardrails, AI Red Teaming and the AI Defense Plane, launched in March 2026.

Key Features

• Prompt Defense: Lakera Guard detects prompt injection and jailbreak attempts in real time, including indirect attacks hidden in documents or web pages. Check Point cites detection rates above 98%, so teams can block manipulation before the model follows hostile instructions.
• Data leakage prevention: Guard scans prompts and outputs for personal and sensitive data before it leaves your application. This helps companies stop customer details, credentials or internal information from appearing in model responses or being pasted into external AI tools.
• Content moderation: Lakera filters toxic, harmful and inappropriate content in both user input and model output. Policies let you decide which categories to flag, which matters for customer-facing chatbots that carry brand and legal risk.
• Malicious link detection: Guard flags suspicious or unknown URLs that appear in prompts or generated answers. This blocks a common route for phishing, where an attacker tricks a model into presenting a harmful link as trusted advice.
• Agent and audio defence: The documentation lists agent behaviour and audio defences alongside the core text guardrails. These extend screening to AI agents that call tools and to voice interfaces, reflecting the shift towards agentic and multimodal applications.
• Custom guardrails and policies: Lakera lets you build custom guardrails, allow lists and deny lists, then group them into policies per project. Security teams can tune strictness for each application instead of applying one blanket rule everywhere.
• Low latency at scale: Lakera claims sub-50 millisecond runtime latency and hundreds of prompts per second. Fast screening matters because every guardrail check adds delay to the user’s chat, and slow checks quickly degrade the product experience.
• Multilingual coverage: Lakera says it monitors attacks in more than 100 languages. That coverage matters because attackers often switch language to slip past filters trained mainly on English text.
• AI red teaming: Lakera Red, now sold as Check Point AI Red Teaming, runs adversarial tests against AI systems before deployment. It covers direct and indirect attacks and offers an expert-led managed service with remediation guidance.

Pros

• Free tier for developers: Lakera offers a free Community plan with 10,000 screening requests a month, according to its documentation. Developers can test real prompt injection protection in a prototype before asking procurement for an Enterprise contract.
• Attack data from Gandalf: Gandalf gives Lakera a large, constantly refreshed stream of real attack attempts. Check Point says the game has generated over 80 million adversarial patterns, which keeps detectors trained on techniques people actually use.
• Simple API integration: Lakera Guard works through one API call that returns a structured verdict with flagged categories. Developers can add screening to an existing LLM app without changing models, since the service is model-agnostic.
• Low false positives: Lakera reports a production false positive rate of 0.01% on its homepage, while Check Point’s announcement cites below 0.5%. Either figure suggests legitimate users are rarely blocked, which keeps support tickets down.
• Backing of Check Point: Check Point ownership gives Lakera the support, sales reach and long-term funding of a large security vendor. Its tools are being folded into the Check Point Infinity Platform, which helps buyers already using that stack.
• Enterprise controls: Lakera Enterprise adds role-based access control, SIEM integration, data retention controls and S3 log export. Requests can carry up to 1MB of context, which suits retrieval-heavy applications with long documents.

Cons

• No public Enterprise pricing: Lakera quotes Enterprise prices only through sales. One pricing tracker reports a median contract of $175,000 a year from five purchases, which, if typical, puts full deployment beyond most small teams.
• Branding in transition: Lakera products are being renamed under Check Point, with Guard becoming AI Guardrails and Red becoming AI Red Teaming. Documentation, contracts and support routes may shift during integration, which complicates long-term planning.
• Self-hosting needs Enterprise: Self-hosted guardrails are available, but reviews say they require an Enterprise contract. The free Community plan runs only on Lakera’s cloud, so teams with strict data residency rules cannot test on-premise for free.
• Added network latency: Each check is a remote API call, and one review estimates 5 to 15 milliseconds of median network overhead on top of detector time. Chains of agent steps that each call Guard can add noticeable delay.
• Verdicts, not enforcement: Guard returns a verdict, and your own code decides whether to block, rewrite or log the request. Teams still need engineering work to turn flags into consistent policy across every AI application they run.
• Limited testing workflow: Reviewers note that Lakera focuses on runtime blocking and is weaker at turning findings into regression tests and CI pipelines. Teams wanting a full evaluation workflow may pair it with a separate testing tool.

Who is Using Lakera?

• Enterprise security teams: Fortune 500 companies use it to govern employee AI use and protect in-house AI applications.
• Product and AI engineers: Developers at firms such as Dropbox add Guard to chatbots and agents to block prompt attacks.
• Regulated businesses: Organisations including Pearson, Hinge Health and Nu cite it for safe, compliant AI launches.
• Check Point customers: Existing Check Point buyers can access the technology through the Infinity Platform and AI Defense Plane.

Pricing

Lakera follows a freemium model, with a free Community plan and quote-based Enterprise plan (official documentation, checked October 2026).

Directories disagree on the details: one review lists a 1,000-request free tier and a per-request Pro plan, while a pricing tracker reports a $175,000 median annual contract. The official documentation shows only Community and Enterprise, so rely on a direct quote, because Check Point packaging may change as products merge.

What Makes Lakera Unique?

Gandalf, Lakera’s public prompt hacking game, feeds its detectors with real attack attempts from a large global player base, a source of training data few rivals can match. That loop, combined with sub-50 millisecond screening across more than 100 languages, made Lakera Guard one of the best-known prompt injection firewalls. Since Check Point’s acquisition, the same technology also sits inside a wider enterprise platform covering employee AI use, agents and red teaming.

Compatibilities and Integrations

• API: A REST API screens prompts and responses and returns flagged categories, working with any model provider.
• Deployment: Cloud SaaS through platform.lakera.ai, plus self-hosted guardrails for Enterprise customers.
• Security tooling: Enterprise plans add SIEM integration and S3 log export, and the technology is joining the Check Point Infinity Platform.
• Developer resources: Documentation includes an integration guide, an API reference and an MCP server for AI coding assistants.

Lakera Tutorials

A first prompt injection check takes only a few minutes on the free plan.

1. Create a free Community account at platform.lakera.ai.
2. Generate an API key and try sample prompts in the playground.
3. Set up a project and choose a policy with the guardrails you need.
4. Send each user prompt and model response to the screening API from your app.
5. Block, log or rewrite flagged requests, then review results in the dashboard.

How We Rated It

Summary

Lakera suits developers and security teams that need to stop prompt injection and data leaks in live AI applications. Its main limitation is cost and change: Enterprise pricing is unpublished and the product is mid-way through rebranding under Check Point. The free tier is generous enough for real testing. Start by running a sample of your own production prompts through the playground to gauge false positives before you commit.