AI infrastructure and security
AI infrastructure built to withstand attacks
AI changes the attack surface: models read untrusted text, agents can take actions and automated attacks run day and night. We design and build infrastructure that assumes exactly that: no open ports, least privilege for people and agents, every action recorded, and rules that block when they cannot be guaranteed.
Updated
When to call us
- You deploy AI agents that have access to systems, code or customer data.
- Your AI application processes outside input: emails, documents, web pages.
- You want to self-host, but without a server reachable from the internet.
- An auditor or customer asks how your AI environment is secured.
Number 1 Prompt injection tops the OWASP Top 10 for LLM applications (2025 edition, the current one), and 'Agent Goal Hijack' tops the OWASP Top 10 for Agentic Applications for 2026. OWASP calls it unclear whether fool-proof prevention exists, and recommends least privilege, human approval for high-risk actions and segregating external content, among other measures.
Not the right fit. We do not sell standalone penetration tests or certification audits; we design and build the environment so that you are well prepared for them.
Pick an attack and see where it stops
Five attacks that happen every day, and the layer that stops each one.
- Scanner hunts a known vulnerability
- Stopped: Zero Trust. As soon as a flaw in widely used software becomes public, bots scan the whole internet within hours for servers that still have it. With Cloudflare Zero Trust the server opens an outbound tunnel itself and no port is open; a request without a valid identity never reaches the application, even when it has not been patched yet. Patching is still needed, but it buys you time.
- Stolen password
- Stopped: Zero Trust. Passwords leak through phishing and reuse. Behind Zero Trust a password alone is not enough: sign-in goes through your identity provider with a second factor, per application, and access can be revoked at once when someone leaves.
- Hidden instruction in a document
- Stopped: Input. An attacker hides a command in a PDF or email that an employee asks the AI to summarise. Because external content stays separated from the instruction, the text is summarised and not carried out. This layer is not watertight, which is why two more follow.
- Manipulated agent looks for secrets
- Stopped: Permissions. Suppose an injection does succeed and the agent is told to look up passwords or keys. The agent only has permissions for its task and cannot reach them.
- Agent wants to deploy or delete something
- Stopped: Approval. An agent, manipulated or simply wrong, prepares an irreversible action. It is only carried out after a human has looked at it and approved it.
Every step is in the audit log, including the blocked ones.
What you get
Zero-trust access
Services are reachable only through a tunnel with identity checks. No ports are open to the internet.
Least-privilege agents
An agent gets exactly the permissions for its task. Destructive actions require a human and a separate key.
Prompt injection defence
Web pages, documents and repository content count as untrusted input: data, not instructions.
Audit logging and secrets management
Every mutating action is attributed and logged. Tokens and keys never appear in code, logs or chat messages.
Hardened containers
Pinned images, no privilege escalation, internal networks and authenticated connections between services.
Application security
Content Security Policy, rate limiting, secure session cookies and device attestation against abuse of public features.
How we approach it
Technology
- Cloudflare Zero Trust
- Cloudflare Access
- Docker
- Coolify
- Linux
- Firebase App Check
- Play Integrity
- DeviceCheck
- CSP
- GitHub Actions
- 01
Threat model
We map what there is to gain, who can reach it and which input is untrusted, including what a model or agent reads.
- 02
Closed by default
We design from no access and open only what is needed. Rules that cannot be guaranteed block.
- 03
Make it provable
Security rules get tests, changes go through review and the log shows who did what.
Proof from our own work
AI platform for a financial services firm: access through Zero Trust with single sign-on, pinned containers without extra privileges and our own security audit of the server.
A private AI platform for a financial services firmOur own agent environment: agents cannot push, merge or change production data, and every action is in an audit log.
Shipping software with AI agents, safely and verifiablyPublic voting in apps with 35,000 active users, protected against double voting with device attestation.
One platform, seven apps: the carnival platformThis website itself: a strict Content-Security-Policy, no trackers, third-party scripts only once you use the contact form, and TLS 1.2 as a minimum.
Frequently asked questions
What is prompt injection?
In a prompt injection an attacker hides instructions in text an AI model reads, such as a web page, email or document. The model may follow those instructions as if they came from the user. The defence is architecture: treat untrusted input as data and limit what the model is allowed to do.
What is Zero Trust, and what exactly does Cloudflare Zero Trust do?
Zero Trust means no request is trusted because of where it comes from: identity first, then access. With Cloudflare Zero Trust your server opens an outbound tunnel to Cloudflare, so no ports need to be open to the internet, and Cloudflare Access checks per application who signs in, with single sign-on and a second factor. Bots that scan the internet for vulnerable software therefore never reach your application. It does not replace patching, and it only protects what really sits behind the tunnel: an admin screen that is also reachable directly on the IP address stays open. That is why we always verify it from the outside.
Why is security different for AI than for ordinary software?
Because a model makes no hard distinction between instructions and data, and because agents act on their own. A mistake in permissions or input handling therefore has consequences sooner than in classic software.
Can AI agents safely access our code and systems?
Yes, if the environment is set up for it. In our own setup agents work in an isolated working copy, cannot publish anything without checks and human approval, and are not allowed to change production data.