Skip to content

Make it say what it shouldn't

AI Disinformation
Testing

We test whether your AI systems can be pushed into stating false, misleading or harmful things in your name: invented facts, fake citations, wrong advice in regulated areas, and answers corrupted by poisoned sources. You learn where the output can be manipulated, how easily, and which controls stop it.

Truth under pressure

Test what it says
under manipulation.

Testing maps to OWASP LLM09 Misinformation, the confabulation and information integrity risks in NIST AI 600-1, and MITRE ATLAS. We define what the system must never say, then attack it through user prompts and through the sources it retrieves, measuring how often each attack succeeds.

Phase 01 · Define the truth

Agree what the system must never say.

A written set of facts, limits and forbidden claims that every test result is judged against.

  • Authoritative sources for your products, policies and prices identified.
  • Regulated topics such as medical, legal or financial advice listed.
  • Brand, safety and legal red lines agreed with their owners.

Phase 02 · Build attack sets

Write the prompts a manipulator would use.

A test corpus of adversarial prompts and poisoned documents built for your system and audience.

  • Leading questions, false premises and persona attacks written.
  • Poisoned documents and web content prepared for retrieval tests.
  • Scenarios based on public cases of AI misuse.

Phase 03 · Manipulate

Push the system into false or harmful answers.

Every successful manipulation recorded with the input, the output and the false claim it produced.

  • Direct prompts that coax out invented facts and fake citations.
  • Indirect attacks through sources the system retrieves or trusts.
  • Long conversations that wear down its guardrails.

Phase 04 · Measure

Find out how often each attack works.

Success rates per attack type across repeated runs, so you know which failures are rare and which are routine.

  • Each attack run many times to account for variable model output.
  • Results scored against the agreed truth set.
  • Failures grouped by root cause: prompt, retrieval or missing control.

Phase 05 · Report & controls

Hand over the evidence and the controls to add.

A report with reproductions and success rates, plus controls ranked by how much false output they stop.

  • Grounding, citation checks and output filters recommended where they fit.
  • Source integrity controls for retrieval pipelines.
  • A regression set your team can rerun after every model change.

Agree what the system must never say.

A written set of facts, limits and forbidden claims that every test result is judged against.

  • Authoritative sources for your products, policies and prices identified.
  • Regulated topics such as medical, legal or financial advice listed.
  • Brand, safety and legal red lines agreed with their owners.

Key benefits

Catch false answers before your customers do.

Failures you can measure

One bad answer can be dismissed as a fluke. We run each attack many times and report how often it works, so you can decide which risks need a control and which you accept.

Protection for your name

When your AI states something false, it speaks for your company. Testing shows where it can be pushed into claims that create legal, regulatory or reputational exposure.

Evidence for AI governance

Results map to NIST AI 600-1, the OWASP Top 10 for LLM Applications and MITRE ATLAS, giving your AI governance group documented, independent testing to show regulators and customers.

What you get

What lands on
your desk.

Built so the product owner can decide on risk and the engineers can build the controls.

The document

The disinformation test report

The truth set, attack methods, every successful manipulation with input and output, success rates by attack type, and controls ranked by the false output they would stop.

The working file

The adversarial test set

The prompts, poisoned documents and expected answers we used, packaged so your team can rerun them after each model, prompt or data change and see whether results hold.

The conversation

The risk review

A session with product, legal and AI engineering. We show the worst outputs, discuss which risks you are willing to accept, and agree the controls to build first.

Client case studies

Real engagements.
Documented outcomes.

YHB | CPAs & Consultants testimonial

In their words

"ForgePath has become a trusted security partner for YHB. Their team understands how a CPA firm operates and has helped us build a security program that’s both practical and sustainable."

Cyrus Kapadia
Chief Technology & Information Officer, YHB | CPAs & Consultants

Trusted by industry leaders.

Before we test

What buyers
ask us.

Straight answers from the practitioners who do the work. If yours is not here, ask us directly.

AI pen testing looks for security breaks such as data leakage and tool abuse. This test looks at the truth of what the system says: whether it can be steered into false, misleading or harmful statements. Many teams need both, and they can share one scope.

Hallucination happens without anyone trying. We measure what a deliberate manipulator can make it say, through prompts and through the sources it trusts, and how often. That tells you which failures need a control.

No. We need test accounts, the system prompt or a description of it, and a list of the sources the system retrieves from. If you want poisoning tests on retrieval, we agree with you where test documents can be placed.

Poisoned documents go only into sources you approve, ideally in staging, and are removed when testing ends. Every item we plant is logged so your team can confirm it is gone.

Ready?

Bring us
the problem.

One conversation with a senior practitioner who has carried this responsibility before. We will tell you what we would do, and whether it is us.