Prompt injection · BlockedAnswer · Unsupported

AI Practice/AI Assurance

Assure · Independent testing

Prove it before anyone asks.

Independent evaluation and red teaming for AI systems, ours or anyone else's. We test them against cases from your own world, score them against a simple baseline, and report plainly what passed, what failed and why.

  • The method behind our public leaderboard
  • Any vendor, any model
  • Tested inside your walls

Why test AI

Confident is not the same as correct.

AI systems fail quietly. Testing them on your own cases is the only way to know how.

Confident and wrong

Fluent answers hide stale, partial and conflicting evidence. Users rarely notice until it matters.

New ways to attack

Chatbots and agents can be tricked into leaking data or taking actions nobody approved.

Buyers now ask for evidence

Boards, regulators and public buyers increasingly ask how an AI system was tested before they rely on it.

Services

Independent testing, before and after go-live.

Built from the method we use on our own products, and run on your cases, inside your environment.

Measure · 3 to 6 weeks

AI evaluations

Test cases built with your own experts, scored against a simple baseline, so you see what the AI adds and where it fails.

  • Cases from your real work, not public benchmarks
  • Behaviours scored one by one
  • A plain report your board can read
Talk to us about this →
Attack · 2 to 4 weeks

Red teaming

We attack chatbots and agents the way a determined user would: data leakage, prompt injection, unsafe actions and harmful answers.

  • Text, document and agent attacks
  • Each finding rated and reproducible
  • Fixes re-tested before sign-off
Talk to us about this →
Verify · before a tender

Product claim review

For vendors: we check that what your product sheets and bids say about AI is true, and help you evidence it.

  • Claims tested against real cases
  • Evidence packs for bids and due diligence
  • Gaps named before a buyer finds them
Talk to us about this →
Select · tender stage

Buyer-side assurance

For buyers: we help write the AI part of a tender and test the shortlisted vendors on your own cases before you sign.

  • AI requirements you can actually test
  • Shortlisted vendors scored side by side
  • A recommendation backed by evidence
Talk to us about this →
Watch · quarterly

Ongoing monitoring

Re-tests after every model or data change, so quality does not drift silently after go-live.

  • Your evaluation set, re-run on a schedule
  • Alerts when a behaviour gets worse
  • A short report every quarter
Talk to us about this →
Govern · ongoing

AI governance as a service

We keep your register of AI systems current, re-test them on a schedule, and report to the board each quarter.

  • One register of every AI system you run
  • Risk, owner and last test for each
  • Board reporting without the busywork
Talk to us about this →

How we test

Built so the AI can fail.

The same method we use on Kybern Prime in public. A test that cannot fail proves nothing.

  1. 01Build the cases

    With your experts, from your real work, including the traps: missing facts, conflicting sources, outdated rules.

  2. 02Set a baseline

    A simple, honest comparison, so you can see what the AI actually adds over doing nothing clever.

  3. 03Score behaviour by behaviour

    Not one accuracy number. Each behaviour that matters is scored on its own.

  4. 04Report plainly

    What passed, what failed and why, with every finding reproducible and mapped to the framework you use.

Behaviours we score, among others

  • Missing information stays unknown
  • Source conflicts are kept, not silently resolved
  • Recognises when the model may not fit
  • Revises when reality changes
  • Abstains instead of guessing
  • Decides when the evidence is enough
259 / 259Reality Challenge cases passed by Kybern Prime, against 74 for a naive baseline, scored in public. See the leaderboard →

Frameworks

Findings your auditors already recognise.

NIST AI Risk Management Framework

Findings mapped to the framework's functions, so they slot into your existing risk process.

OWASP guidance for AI applications

Red-team findings organised against the most common weaknesses in AI and agent applications.

Your sector's rules

Banking, public sector and defence requirements, and your own internal policies.

Questions

Questions about assurance.

Something else on your mind? Ask us directly.

Can you test a system you did not build?

Yes. We test AI from any vendor, open or commercial models, chatbots, agents and document AI. Independence is the point.

Can testing happen inside our environment?

Yes. We can run evaluations and red teaming on your premises and fully offline, so your data and the system under test never leave.

Will you publish our results?

No, not without your written permission. We publish our own results; yours belong to you.

We are a vendor. Can you help us prepare for a buyer's test?

Yes. A product claim review checks what your materials say about AI against real cases and builds the evidence pack before a buyer asks.

AI Assurance

Tell us what you need to trust.

A system you run, a product you sell or a vendor you are choosing. We will suggest how to test it.

  • A conversation with someone who does the work, not a script.
  • A suggested first step, sized to your problem.
  • A clear view of scope, cost and how success is measured.

Prefer to talk now? Call +91 98150 23001 or write to reach@requisitetech.com.

We reply within one working day. We use your details only to respond to this request. See our privacy notice.