Verdict: Lakera Guard sits between users and your LLM, screening every input and output for prompt injection, jailbreaks, and data leakage before they reach the model — at sub-50ms latency and a reported 98%+ detection rate. It’s a different problem than deepfake detection: this protects an AI application itself, not media authenticity.
Quick Facts
| Protects against | Prompt injection, jailbreak attempts, PII exposure and data leakage, toxic content |
| How it works | Real-time runtime layer — screens every input/output before it reaches or leaves the LLM |
| Reported performance | 98%+ detection rate, sub-50ms latency, false positive rate reported between 0.01% and 0.5% depending on source |
| Coverage | Multimodal, model-agnostic, 100+ languages |
| Made by | Lakera (founded 2021, Zurich) — acquired by Check Point in September 2025 |
| Free tier | Yes — generous enough for evaluation and small deployments |
| Paid pricing | Not published; user reports put meaningful production usage around $500+/month, scaling to a few thousand/month at 1–5M requests |
Last tested: August 2026
How It Actually Works
Lakera Guard is a runtime security layer, not a one-time scan. Every message going into your LLM, and every response coming back out, passes through Guard’s detection engine first — if something looks like a prompt injection, jailbreak attempt, or an attempt to exfiltrate sensitive data, it gets flagged or blocked before it reaches the model or the user. The detection models are trained partly on data from Gandalf, Lakera’s own educational prompt-injection game, which has collected over 80 million adversarial prompts from more than a million players — a genuinely unusual threat-intelligence pipeline most competitors don’t have.
Where It Falls Short
Like most enterprise-grade AI security tools, pricing isn’t published — you’ll need a sales conversation to get a real number, and independent reports suggest meaningful production usage lands well above what an indie developer or small side project would want to pay. If your LLM app is small-scale or pre-revenue, the free tier may be worth testing before assuming you need the paid product.
Our Team’s Take
We don’t run a customer-facing LLM application ourselves, so we can’t offer firsthand production experience with Guard’s runtime layer. What stood out across independent reviews is the Gandalf threat-intelligence angle — most competitors rely on internal red-teaming alone, while Lakera has an unusually large, continuously growing adversarial dataset feeding its detection models.
(Note to editor: this section reflects desk research, not hands-on testing — if your team ever builds a customer-facing AI tool, a real Guard integration test would make this section far stronger.)
When Not to Use Lakera Guard
- You’re an indie developer on a tight budget. Reported production costs start around $500/month — a self-hosted open-source alternative may fit better early on.
- You need transparent, self-serve pricing. Like Reality Defender and Sensity, getting a real number means talking to sales.
- You need media (deepfake) detection, not LLM security. Guard protects your AI application from malicious inputs — it doesn’t detect fake video, audio, or images.
How It Compares
Lakera Guard solves a fundamentally different problem than Reality Defender and Sensity AI — it defends an LLM application from malicious users, not the public from deepfake media. Within its own category, its main differentiator is the Gandalf-sourced threat intelligence and its open-source PINT benchmark, which independent reviewers cite as an unusually strong credibility signal for a security vendor.
Our Testing Method
We cross-checked Lakera’s product documentation, the Check Point acquisition details, and independent security-tool comparisons published through mid-2026, without a hands-on production integration test.
Last updated: August 2026
What is Lakera Guard?
A real-time security layer for LLM applications that screens input and output for prompt injection, jailbreaks, and data leakage.
Is Lakera Guard free?
There’s a free tier generous enough for evaluation and small deployments; production-scale usage requires a custom quote.
What is Gandalf?
Lakera’s educational prompt-injection game, used by over a million players, that feeds real adversarial-prompt data into Guard’s detection models.
Who owns Lakera now?
Check Point acquired Lakera in September 2025 in a deal reported around $300 million.
