Skip to content
Civware

AI tested against misuse,before someone else tries

We test your AI systems the way someone trying to misuse them would: hidden instructions, attempts to pull out data, and requests it should refuse. You get findings ranked by risk, and help fixing them.

Group
Run
First step
A test of one AI system
You decide
Which fixes to make first
Works with
Assistants, agents and AI features

When it fits

Worth doing before launch, and again whenever a system gains new tools or data.

  1. 01

    An AI system is about to reach customers or connect to sensitive data.

  2. 02

    Your security team needs evidence that AI risks were tested, not assumed.

  3. 03

    An agent is getting new tools or wider access.

What we build

  1. 01Tests for prompt injection: text hidden in emails, files or web pages that tries to redirect the AI.
  2. 02Tests for data leaks: attempts to get the AI to reveal what it shouldn't.
  3. 03Tests for unsafe actions: attempts to make an agent do more than its task.
  4. 04A findings report, ranked by risk, with a plain description of each finding.
  5. 05A retest after fixes, so you know they worked.

How it works

Each attempt is recorded, and each finding is ranked by how likely it is and how much harm it could do.

  1. 01We try hidden instructions, data leaks, unsafe actions and misuse.
  2. 02Findings are ranked High, Medium or Low, so the riskiest is fixed first.
  3. 03Fixes are retested before the report is closed.
Findings, ranked by riskHidden instructions, data leaks, unsafe actions and misuse are tried against a test copy of your AI system. Three findings are ranked: following hidden text in files is high risk and fixed first, showing other users data is medium, and long replies raising costs is low and monitored.FIG. 01FINDINGS, RANKED BY RISKWHAT WE TRYhidden instructionsdata leaksunsafe actionsmisuseYOUR AI SYSTEMTEST COPYFINDINGRISKFollows hidden text in filesfix first: treat files as dataHighCan show other users' datafix next: tighten accessMediumLong replies raise costsmonitor: set a limitLow
Fig. 01 Findings, ranked by risk. The highest-risk finding is fixed first, then retested.

Safeguards

Security testing has to be safe itself. These are the rules we work to.

Testing that affects real users
Tests run against a test copy of the system, or in agreed windows, with your security team informed.
Findings reaching the wrong people
Reports go only to the people you name, and test data is deleted at the end.
Fixes that create new problems
Every fix is retested, along with the cases that passed before.
A one-off test going stale
We recommend a retest when the system gains new tools, data or users.

How to start

First engagement · Security test

One AI system, tested and ranked

It delivers:

  • Testing for prompt injection, data leaks and unsafe actions
  • A findings report, ranked by risk
  • Recommended fixes, in plain words
  • A retest once the fixes are in
Price
Fixed, agreed before we start
Length
Usually 2 to 3 weeks
Decision pointYou decide which fixes to make, and when. We retest the ones you choose.

Questions

Is this the same as a penetration test?

It's related but different. A penetration test checks your systems' defenses; this checks how the AI itself can be misused. The two work well together.

What is prompt injection, in plain words?

Text hidden in something the AI reads, such as an email or a file, that tries to give it new instructions. It is one of the main risks for AI that reads outside content.

Can you test a system we bought?

Often, yes, with the vendor's permission. We'll check before we start.

Will you fix the problems?

We'll recommend fixes and can make them in systems we built. For others, we'll work with whoever maintains them.

Tell us one job you'd like AI to do

We'll tell you honestly whether AI is the right fit, and what a sensible first step looks like.

One job is enough to start.

We use your details to reply to you, as described in our privacy policy.