The vendor tested the model. Nobody tested the system you built around it.

An AI app includes the model, prompts, data pipelines, connectors, and permissions. Attacks often arrive as ordinary inputs—documents, tickets, or web pages—and exploit meaning, not code.

These 5 quick reads are for the teams building and deploying AI.
See the 5 reads

3 questions to answer before the next release

Every AI application has to answer three questions before it ships: what a hostile input can make it do, what its tools can reach, and where the guardrail actually runs.

What can a hostile document make your app do?

Indirect prompt injection rides in on retrieved content. A predefined prompt suite never sends it. Adaptive, application-specific attack paths do, and they change every time you ship.


What is the blast radius of your agent's tool set?

A calendar tool is harmless. Add CRM read and an email connector and the same agent can exfiltrate a customer list on request. Risk is the context, not the component.


Where does your guardrail run, and who owns it?

String-matching rules fail on meaning. Per-app regex does not scale. Enforcement belongs at runtime, before the action, as one policy layer across apps, agents and people.


Start with the Red Teaming whitepaper

Running in production, at latency you can live with

Both enforce on live, customer-facing AI traffic, in front of LLM-powered applications and regulated workflows.

*Previously known as Lakera


"Dropbox uses AI Agent Security as our security solution to help safeguard our LLM-powered applications, secure and protect user data, and uphold the reliability and trustworthiness of our intelligent features."

- Dropbox Engineering



"We've chosen Check Point* to secure our enterprise GenAI deployment across our regulated banking environment.This partnership enables us to safely innovate with AI in money transfers, financial services, and customer support. Check Point's accuracy, low latency, seamless integration, scalability and  support for Portuguese and Spanish are essential for our global operations, especially in markets with sophisticated fraud attempts."
- Nu


Anchor: resources

Five reads. One idea each.

Under 15 minutes apiece. Start anywhere.

Bring an app. Watch it get tested.

 30-minute session with an AI security engineer: how a red team run against an LLM app like yours is scoped, what a finding looks like, and how runtime policy is enforced without latency you would notice.
Talk to an AI Security Engineer

Every control is doing its job. None of them are watching behavior.

White paper, 19 pages, 13 min. What WAF, AppSec, DLP and IAM each see and miss, the four new data paths, MCP as the agent execution surface, and one policy layer across apps, agents and people.

Your AI passed the test. Production is a different exam.

White paper, 14 pages, 10 min. Why predefined prompt suites miss adaptive, application-specific attack paths, the four-stage Start / Shape / Act / Expose model, and what to evaluate across design...

AI stopped asking permission. Your security stack did not notice.

Executive guide, 22 pages, 15 min. The one to send your CISO: five documented incidents, where the risk lives, and the first five moves. Plain language, sourced, quotable.

Shadow AI does not look like shadow IT. It looks like getting work done.

White paper, 17 pages, 12 min. Coding assistants with the surrounding code in the prompt, personal accounts on work machines, MCP-connected workflows: where unmanaged AI use hides and what point-...

Your agents are approved. Their actions are not.

White paper, 17 pages, 11 min. The eight components an agent is assembled from, where each becomes an attack surface, and how a runtime risk score turns Allow, Require approval and Block into a p...