AI Gatewayby SecureAxis

Early access · working prototype

AI answers are the new phishing email.

AI Gateway sits between your people, your apps and AI model APIs. It checks every prompt going out and every answer coming back for phishing links, hidden instructions, leaked secrets and malicious install commands.

Works with Claude and GPT through one OpenAI-compatible endpoint. Runs in your own environment.

Illustration with made-up names.

Your email filter never sees this traffic

AI assistants have become a new way to deliver phishing and malware. These are the attacks the gateway is built to catch.

Invisible text

Hidden-character smuggling

Instructions or lure text hidden in invisible Unicode characters. Filters see harmless words while the model reads the hidden payload. The gateway decodes these first, and smuggled text is a hard block.

Read the BleepingComputer report ↗
Poisoned answers

Lookalike installers

A model is steered into recommending a lookalike download site and a one-line terminal command. The gateway flags pipe-to-shell installs and installer links on unverified domains, then visits the link in a sandbox.

Agent config

Poisoned instruction files

An agent's instruction file (such as SKILL.md, CLAUDE.md or .cursorrules) quietly tells it to download something every session. A file-scan endpoint blocks run-on-load, “don't tell the user” and credential-stealing instructions.

Links

Phishing links in answers

Link shorteners, lookalike characters, brand names on the wrong domain and links buried in PDFs. Shorteners are expanded, PDFs inspected and suspicious links visited in a sandbox.

Injection

Prompt injection

“Ignore previous instructions”, fake system tags and requests to send data elsewhere. Scored by rules first, then by an independent judge model.

Leakage

Secrets in prompts

API keys, passwords, tokens and card numbers pasted into a chat. They are redacted before anything leaves your network.

How it works

Rule-based checks run in order, each on the text the previous one cleaned. Reputation lookups and the judge model then run side by side. Flagged links and files go to the sandbox, and what it finds feeds a blocklist that later checks use.

  1. 01NormaliseInvisible and lookalike characters decoded
  2. 02SecretsKeys, tokens, passwords and card numbers redacted
  3. 03InjectionInstruction overrides and fake system tags scored
  4. 04CommandsPipe-to-shell and risky install commands flagged
  5. 05LinksDomains checked against brands, shorteners expanded

Then, side by side

ReputationPublic threat feeds such as VirusTotal and URLhaus, plus checks on recommended GitHub, PyPI and npm packages
Judge modelA small, independent model scores phishing, injection, data theft and unsafe downloads
Verdict and auditAllowed, cleaned or blocked, with the evidence attached and a JSON record for your SIEM

Don't trust the answer. Detonate it.

Most AI guardrails ask a second model whether an answer looks safe. This gateway goes and checks.

Suspicious links and files are opened in a disposable sandbox, and the verdict comes from what actually happened, not from a model's opinion. The gateway never runs anything itself; it hands the work to an isolated backend.

BackendChecksWhat comes back
Built-in browser sandboxLinksA fresh browser for every visit: screenshot, redirect chain, hosts contacted, downloads attempted, password forms and where they send data
urlscan.ioLinksScreenshot, landing page, hosts contacted, brand verdict
VirusTotalFilesDetections and behaviour: processes, dropped files, network activity
CAPEv2, self-hostedWindows filesFull virtual-machine detonation
  • A page that names a brand and asks for a password on a domain that isn't that brand's is malicious.
  • A password form that sends data to another site is malicious.
  • Download attempts and redirects to sign-in pages are suspicious.
  • Confirmed bad domains and file hashes go on a blocklist for 30 days, so the next chat that tries the same link is refused straight away.

One line for developers

The gateway speaks the OpenAI API, streaming included. Point your existing client at it and keep your code.

from openai import OpenAI

client = OpenAI(
    base_url="https://ai-gateway.your-company.internal/v1",  # was api.openai.com
    api_key="gateway-key",  # model API keys stay in the gateway
)
reply = client.chat.completions.create(
    model="claude-sonnet",  # or a GPT model; routed by name
    messages=[{"role": "user", "content": "Which VPN client should I install?"}],
)
  • /v1/chat/completions – OpenAI-compatible, with streaming.
  • /v1/scan/file – checks agent instruction files before they are loaded.
  • /audit – JSON-lines audit trail for SIEM pickup.
  • Routes to Anthropic (Claude) and OpenAI (GPT) by model name. Azure OpenAI routing is built in.
  • Your model API keys live only in the gateway. People and apps never see them.

For staff and security teams

A chat that shows its working

Looks like any AI assistant, but every reply carries a verdict. An inspection panel shows what was checked and why, and can reveal hidden characters.

Check a link or file

Paste a URL or drop a file and get a sandbox report back, with a screenshot.

A console for the SOC

An audit stream, judge model cost and threat-feed counters, with per-team policy kept in a simple config file.

How it decides

The design choices behind the verdicts, in plain words.

  • Rules first, models second. Fixed rules run first because a model can be talked out of a decision and a rule can't. The judge model catches what rules miss, such as social-engineering wording.
  • The judge is independent. It comes from a different vendor than the model doing the work, uses a fixed prompt, and its output never flows back into the working model's prompt.
  • Evidence beats opinion. A second model can be fooled by the same trick as the first, so link and file verdicts rely on threat feeds and sandbox results.
  • Clean the text before anything else. Hidden characters are decoded first, so no later check, including the judge, is fooled by them.
  • Fail closed where it matters. If a check times out, an ordinary answer still goes through. An answer containing a link, installer or shell command does not.
  • Allowlist installers. Downloads are trusted only from a maintained list of official domains. Everything else is treated as unverified.

Where it stands

A working prototype with 164 automated tests, checked against real threat feeds. It is not yet ready for production use.

  1. Sandbox vs. a fake sign-in pageRated malicious in 2.7 seconds.
  2. End to end: short link → phishing page → blocklistThe next chat that tried the same link was refused.
  3. Red-team replay with live model APIsAll attack groups blocked.
  4. Lookalike domain and test malware fileA homoglyph domain was flagged, and the industry-standard EICAR test file was caught by its hash.
  5. Smuggled text and poisoned filesA hidden-character prompt, a keyword-split lure, a poisoned reply and a poisoned SKILL.md: blocked, cleaned, annotated and blocked.

Questions

Do we need to change our apps?

No code changes beyond the address. Apps that use an OpenAI-compatible client change their base_url to the gateway and keep working, streaming included.

Which AI models does it work with?

Claude (Anthropic) and GPT (OpenAI), chosen by model name. Azure OpenAI routing is built in. Your model API keys stay in the gateway, so people and apps never handle them.

Where does it run?

In your own environment. It ships as a Docker image with a compose file. Prompts and answers go from the gateway straight to the model provider you choose.

Does it slow replies down?

The rule checks are quick. The judge model and reputation lookups add some time, which is why the judge is a small, fast model. Link and file detonation feeds the blocklist, so a link confirmed as bad is refused immediately the next time anyone meets it.

What happens if a check fails or times out?

An ordinary answer still goes through. An answer that contains a link, an installer or a shell command does not, so a failure can't be used to sneak one past.

Is it ready for production?

Not yet. It is a working prototype with 164 automated tests. The sandbox still needs moving to an isolated network, sign-in needs finishing and the blocklist needs a database. Early access is for teams who want to try it and shape it.

Request early access

For security teams and developers who want to try the gateway and help shape it. We'll reply by email.

Also interested in (optional)

We reply by email from hello@secureaxis.net. No newsletter. Prefer email? Write to hello@secureaxis.net.