Skip to content

Attack your AI first

AI Pen
Testing

We attack your AI applications the way real adversaries do: prompt injection, jailbreaks, data leakage through prompts and retrieval, and abuse of the tools and plugins your models can call. Testing maps to the OWASP Top 10 for LLM Applications and MITRE ATLAS, and every finding comes with a working reproduction.

Model, prompts, tools

What your model
can be talked into.

Testing covers the OWASP Top 10 for LLM Applications and maps to MITRE ATLAS techniques. We combine manual adversarial prompting with tools such as PyRIT and garak, and test the APIs around the model with Burp Suite. Every finding is reproduced by hand before it reaches the report.

Phase 01 · Scope & architecture

Map what the model can see and do.

A diagram of the model, its prompts, data sources and tools, with the trust boundaries attackers will target.

  • System prompts, retrieval sources and connected tools inventoried.
  • User roles defined, with what each should be allowed to reach.
  • Abuse cases agreed with product and legal owners.

Phase 02 · Prompt injection

Make the model ignore its instructions.

Direct and indirect injection attempts that worked, with the exact inputs and the behavior they caused.

  • Direct injection and jailbreak techniques against the system prompt.
  • Indirect injection hidden in documents, web pages and email the model reads.
  • System prompt extraction attempts.

Phase 03 · Data leakage

Pull out data the model should never return.

Sensitive data, secrets or other users' content the model disclosed, with the prompt that got it.

  • Retrieval permissions tested across users and tenants.
  • Disclosure of embedded secrets and training data tested.
  • Vector store and embedding weaknesses examined.

Phase 04 · Tool & plugin abuse

Turn the model's tools against you.

Actions an attacker caused the model to take, such as sending data out or changing records.

  • Excessive agency tested: tools called without proper checks.
  • Model output passed to downstream systems tested for injection.
  • Plugin and API permissions reviewed for least privilege.

Phase 05 · Report & fixes

Hand over the reproductions and the guardrails to add.

A report with working reproductions mapped to OWASP and ATLAS, and fixes your engineers can build.

  • Each finding ranked by impact and ease of exploitation.
  • Fixes at the prompt, retrieval, tool and application layers.
  • Retest of fixes available once they ship.

Map what the model can see and do.

A diagram of the model, its prompts, data sources and tools, with the trust boundaries attackers will target.

  • System prompts, retrieval sources and connected tools inventoried.
  • User roles defined, with what each should be allowed to reach.
  • Abuse cases agreed with product and legal owners.

Key benefits

Find the prompt that breaks it, before users do.

Tested the way attackers work

Canned prompt lists catch the obvious. A senior tester adapts to how your model responds, chains techniques together, and finds the injection that works on your system specifically.

Mapped to OWASP and ATLAS

Findings reference the OWASP Top 10 for LLM Applications and MITRE ATLAS, so your AI governance team, auditors and customers can see exactly which risks were tested.

Fixes beyond the prompt

A better system prompt rarely fixes injection on its own. We recommend controls at the retrieval, tool permission and application layers, where a fix holds up against the next variation.

What you get

What lands on
your desk.

Each piece includes the inputs that worked, so your engineers can reproduce them and confirm the fix.

The document

The AI pen test report

Scope and architecture, findings mapped to the OWASP Top 10 for LLM Applications and MITRE ATLAS, each with the exact prompts, the model's responses, impact and the recommended fix.

The working file

The prompt reproduction set

Every successful attack as a reproducible input with expected and actual behavior. Add it to your regression tests to check each fix still holds after a model or prompt change.

The conversation

The AI engineering readout

A session with your AI engineers and product owners. We replay the most serious attacks live, walk through the fixes, and answer questions from the people who will build them.

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.

AI testing questions

Asked before
every AI test.

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

Your application. Foundation models from commercial providers are the provider's to test. We test how you use them: your prompts, retrieval, tools, permissions and the code around the model, which is where your own exposure sits.

Test accounts at each user role, documentation of the system prompt and connected tools, and ideally a staging environment. Access to source code or prompt configuration lets us test more deeply in the same engagement.

We test in staging where possible. In production we use dedicated test accounts, avoid actions that change real records, and agree limits on token spend and request rates before we start.

Yes. Both call for evaluating AI systems for risk and acting on what you find. The report gives you dated, independent evidence of adversarial testing and the remediation that followed.

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.