Lakera is an AI security company that protects generative AI applications, AI agents and employee AI use from prompt attacks and data leakage. It is built for enterprise teams putting large language models and agents into production, where traditional security controls cannot see what goes into and comes out of a model.
The aim is to let those teams ship AI features and agents without handing attackers a new way in through the prompt, the model’s output or the tools an agent calls.
At A Glance
| Vendor | Lakera (acquisition by Check Point announced September 2025) |
|---|---|
| Category | AI Safety and Guardrails, AI Model Security |
| Identities covered | AI agents, MCP-connected tools, employee AI use |
| Deployment | SaaS or self-hosted (Guard) |
| Agentic AI | Agent discovery, runtime enforcement of agent actions, MCP tool protection |
| Customers | Dropbox, Pearson |
| Founded | 2021 |
| Market classification | AI Safety & Guardrails; also AI Model Security |
What Lakera Does
Generative AI opens new routes for attack. A crafted prompt can override a model’s instructions, a jailbreak can push it into harmful output, sensitive data can leak through prompts and responses, and agents can act on instructions hidden in the tools they call. Lakera addresses these risks at three points: at runtime, by screening every model interaction; before release, by testing AI applications against the same attacks; and in the workforce, by governing which AI services employees use and what data they share with them. Check Point now offers the runtime and red teaming products as Check Point AI Agent Security and Check Point AI Red Teaming.
Products
AI Agent Security (Lakera Guard)
Lakera Guard sits in front of AI applications and agents and checks each prompt, response and agent step. It blocks prompt injection, jailbreaks and adversarial instructions before they reach the model, detects PII and other sensitive data, moderates content and catches malicious links. The prompt attack detector also screens tool responses and tool descriptions, so a compromised tool cannot slip instructions to an agent. For agent deployments it also discovers agents and MCP-connected tools, including agents teams “did not explicitly build or register”, and enforces policy over what those agents can access, call and do. Guard needs no changes to models or prompts, and covers more than 100 languages with a reported runtime latency below 50 ms.
AI Red Teaming
AI Red Teaming gives security and AI teams a continuous way to evaluate, scan and attack their own applications and agents. Tests cover prompt injection, jailbreaks, data leakage, unauthorised actions and harmful or policy-breaking output, and can be tailored to the architecture and usage of a specific system. Regression runs after a model update or system change show whether protections still hold. An expert-led red teaming service is also available.
Workforce AI Security
Workforce AI Security is for organisations whose staff already use public AI services at work. It discovers shadow AI use, governs which services employees can reach and protects the data people put into prompts.
Gandalf
Gandalf is Lakera’s adversarial AI game and network, in which players try to trick a model into revealing a secret. The attempts form a large body of attack data, with a reported 80 million adversarial patterns collected.
How It Works
Guard works on a simple request model: the application or AI gateway makes a call to the Guard API for each user interaction or agent step, and acts on the result.
| Component | What it does |
|---|---|
| Projects | A project is the unit of protection, usually one per application, with its own policy and mode. |
| Policies | A policy sets which detectors apply to an application, chosen from a catalogue or written as a custom policy. |
| Modes | A project either blocks flagged interactions or logs them for investigation. |
| SaaS deployment | A hosted service with the dashboard, autoscaling and daily base model updates. |
| Self-hosted deployment | The customer runs Guard, with base models updated at least every stable release. |
Guard is model agnostic, so it works with hosted model providers, open-source models and custom or fine-tuned models, and can sit behind an AI gateway or be called directly from an application. In SaaS deployments logs can go to a SIEM; self-hosted Guard writes structured logs for third-party logging and observability stacks.
Identities And Systems Covered
| AI agents | Agents across the environment, including unregistered ones, and the actions they take |
|---|---|
| MCP-connected tools | Tool responses and tool descriptions, screened for indirect injection |
| Employees | Employee use of AI services and shadow AI |
| Agent platforms | Amazon Bedrock and AgentCore, Google Cloud, Microsoft Copilot Studio, Salesforce Agentforce, n8n and Relevance AI |
Company Background
| Founded | 2021, by David Haber, Dr Mateo Rojas-Carulla and Dr Matthias Kraft |
|---|---|
| Headquarters | Dual headquarters in Zurich and San Francisco, with AI R&D centres in both cities |
| Ownership | Check Point announced an agreement to acquire Lakera on 16 September 2025 |
Note: This overview is compiled from Lakera’s publicly available information as of 3 November 2026. It is provided for information only. NHI Mgmt Group does not endorse any vendor, and nothing on this page represents the opinions of NHI Mgmt Group.