Vigilant Cybersecurity

AI testing for the attacks most programs don't cover.

If your AI reads text you do not control and can take actions, it can be tricked. We test prompt injection, data leaks and agents that can do too much.

Talk to a person

Member of Anthropic's Cyber Verification Program

AI / LLM Security Testing

What we test.

If your model can read untrusted input and take actions, you have an attack surface most testing programs have no coverage for. We test prompt injection, tool and agent abuse, training data exposure, and the guardrails you believe are holding.

01

Direct and indirect prompt injection

We try to take over your AI with text typed in, or hidden in a web page, email or file it reads.

02

Tool, plugin, and agent action abuse

We try to make the AI use its tools in ways you never meant, like sending email or deleting data.

03

System prompt and training data extraction

We try to get the AI to reveal its hidden instructions or private data.

04

Guardrail and content filter bypass

We test whether the safety rules you set can be talked around.

05

Excessive agency and permission scope

We check whether the AI has more access than its job needs.

06

RAG pipeline and vector store poisoning

We test whether bad documents can be slipped into what the AI searches, to change its answers.

What we look for first.

These are the problems that turn up again and again in this kind of system. We check for them first, then go wherever your setup leads.

Hidden orders in content

Text in a document or web page that takes control of the AI.

Leaked instructions

The AI repeats its private instructions when asked the right way.

Agents with too much power

An AI that can act on accounts or data it never needs.

One customer reaching another

Search results that pull in documents from other customers.

How the test runs.

  1. 1

    Agree on scope

    We list the AI features, tools and data sources in scope.

  2. 2

    Map what it can do

    We learn what the AI can read, which tools it can use and who can talk to it.

  3. 3

    Try to trick it

    We test by hand, with direct and hidden attacks, and test the app around it too.

  4. 4

    Report and retest

    Every finding comes with proof, the risk in plain English and how to fix it. With the Complete package, we test again after your fixes.

What we need from you

  • Access to the AI features, as each user role
  • A list of the tools and data sources the AI can use
  • A test environment if the AI can take real actions
  • A contact we can reach during testing

Getting this for an audit?

Every finding is tied to the rule it affects. See what auditors look for:

Everything your auditor will ask for.

One fixed price covers all of it.

One price, locked when you sign.

We confirm the scope with you first. After you sign, the bill never grows.

Your priceLocked
What we test
Website and API
Start
Within 1 week
Retest
In one price
Surprise bills
Never

The letter your auditor asks for.

Signed at the end of the Complete package, ready to send to whoever asked you for a pentest.

We test again until it is fixed.

With the Complete package, retests are part of the price. Not a second bill.

  • Admin page open to the internetFixed
  • Old software with known holesFixed
  • Weak password rulesFixed

Big problems reach you the same day.

Not three weeks later in a report.

Serious problem foundTodayHere is what to do right now.

Real people do the testing.

Senior testers check every finding by hand. A scanner is where we start, not where we stop.

  • OSCP
  • OSCE
  • GPEN
  • GWAPT
  • CISSP
  • CISA
  • CEH
  • CRTO
  • AWS Security
  • Azure Security

A report you can actually read.

Plain English, with the fixes that matter most at the top.

Fix first
Fix next
Fix later

AI security testing questions, answered.

01What is prompt injection?

It is when text the AI reads gives it new orders. It can be typed in directly, or hidden in a web page, email or file the AI is asked to read.

02Do you test AI agents?

Yes. If your AI can take actions, like sending messages or changing records, that is where we focus most.

03We use a model from another company. Do you test the model?

We test how your product uses the model: your prompts, tools, data and the app around it. That is where most of the risk is, and it is the part you control.

04Does it cover the OWASP Top 10 for LLMs?

Yes. We test the risks on that list by hand, plus the ones that come from how your own product works.

05How long does it take?

We can start within a week. Most tests then take one to three weeks, depending on size. If you have an audit date, tell us on the scope call and we will plan around it.

Your proposal is 90 seconds away.

Type your website. Scope, timeline and a fixed price, in about 90 seconds. Free, and no sales call.

By getting a proposal, you agree to our Terms and Privacy Policy.

  • No surprise bills
  • Start within 1 week
  • Retest and attestation letter in one price

Rather talk to a person? Book a call