Skip to content

AI · LLM Security Testing

Find the prompt injection before a user does

An AI feature takes user input and acts on it, and your last pentest never looked at that path. We test the model endpoints, the prompts behind them, and what the output touches, mapped to the OWASP LLM Top 10.

How it’s delivered
  1. 01MapWe list every model input and what each model can reach.
  2. 02TestManual testing of each path, with critical findings raised the day they're demonstrated.
  3. 03ReportFindings mapped to OWASP LLM Top 10 categories, with reproduction steps.
  4. 04RetestFixes verified against the original attack, and the report updated.

Engineering-led cybersecurity and quality engineering since 2020, delivered by 100+ security & quality engineers on platforms we build and run ourselves.

A prompt injection turns your own product into the attacker's tool: it leaks other customers' data, ignores its instructions, or passes hostile output to systems that trust it. Enterprise buyers now ask who tested your AI features.

See a redacted sample report

The structure, depth, and remediation detail your team will receive, with client identity and evidence removed. Verify your email and read a real, redacted report before any sales conversation.

Get the sample report

Coverage against the OWASP LLM Top 10

Each category, and how a test of your live AI features covers it. Two sit outside what an endpoint test can see, and one belongs to quality engineering.

  • LLM01:2025Prompt InjectionDirect injection through every input that reaches the model, and indirect injection through files and content the model reads.Tested
  • LLM02:2025Sensitive Information DisclosureAttempts to pull other users' data, secrets, or internal context out through model responses.Tested
  • LLM03:2025Supply ChainModel sourcing and third-party components can't be judged from the endpoint. The report lists what we could not see.Outside this test
  • LLM04:2025Data and Model PoisoningThis lives in your training and fine-tuning pipeline, which a live-endpoint test does not reach.Outside this test
  • LLM05:2025Improper Output HandlingWhat happens when model output lands in a page, a query, a shell, or another API without being treated as untrusted.Tested
  • LLM06:2025Excessive AgencyTested where the model can call tools or act. Systems built around agents get the agentic assessment instead.Where exposed
  • LLM07:2025System Prompt LeakageWhether the system prompt, and anything sensitive placed in it, can be recovered by a user.Tested
  • LLM08:2025Vector and Embedding WeaknessesWhere retrieval is exposed, whether one user can get documents or embeddings that belong to another.Where exposed
  • LLM09:2025MisinformationA quality question more than a security one. Our QE practice validates AI output against product intent.Outside this test
  • LLM10:2025Unbounded ConsumptionRate limits, request size, and cost abuse on the model endpoints, the same checks we run on any API.Tested

Category names quoted from the OWASP Top 10 for LLM Applications 2025.

What the assessment covers

Model inputs, mapped first

Every path where user input, uploaded files, or retrieved content reaches a model, including the ones added since your last release.

Injection and extraction

Direct and indirect prompt injection, system-prompt recovery, and attempts to pull data the user should never see.

Output and downstream trust

Where model output flows into pages, queries, or other services, and whether anything downstream treats it as safe.

The API around the model

Authentication, authorization, rate limits, and upload handling on the endpoints that serve the model. This is where our published finding chained.

Tools & standards

Framework
OWASP Top 10 for LLM Applications 2025
Method
Manual testing aligned to OWASP, PTES, and NIST SP 800-115, the same as our other application work

What you receive

  • Findings mapped to the OWASP LLM Top 10, each with reproduction steps
  • A list of the model inputs we found, including any you didn't know about
  • What the test could not reach, stated plainly
  • A retest-verified final report

Evidence

A real finding, readable in full

An AI rewrite endpoint placed user input straight into the model prompt. We demonstrated the injection and two file-upload bypasses; all three were fixed and verified clean on retest in June 2026.

Read the case study →

The finding as reported

The prompt-injection finding is reproduced from the redacted report, with severity, reproduction steps, and remediation guidance.

Read the finding →

Who this is for

  • Product teams shipping chat, summarization, or rewrite features built on an LLM
  • SaaS companies whose enterprise customers ask how their AI features were tested
  • Teams adding retrieval (RAG) over customer documents

Common questions

Do you need access to the model itself?

No. We test through the same interfaces your users and attackers reach. Risks that live in model sourcing or the training pipeline are outside that test, and the report says so.

Which model providers do you work with?

The assessment targets your application and its endpoints, so the underlying model provider doesn't change the method.

Is retesting included?

Yes. Fixes are verified against the original attack and the report is updated.

Scope an LLM security test

Tell us which AI features are live and what they can reach.