AI Systems

Test what your AI can be made to do.

We attack models, agents, retrieval pipelines and the tools they can call. Every exploit ships with the input that triggered it.

From prompt to consequence.

Six steps from the system map to the retest of your guardrails.

  1. Map the boundaries

    Models, tools, retrieval and identity are drawn as one system with its data boundaries.

  2. Inject and leak

    Prompt injection and data-leakage attacks run against the assistant as users see it.

  3. Misuse the tools

    We make the agent call tools it should not, with data it should not have.

  4. Poison retrieval

    Planted documents show what a hostile corpus can make the system say or do.

  5. Measure the impact

    Each finding carries a reproducible prompt and its impact per data boundary.

  6. Harden and retest

    Guardrail recommendations are tested again after you apply them.

What this engagement covers

When you need it

  • Before exposing an assistant to customers
  • After adding tool access
  • For AI governance evidence

What we test

  • Prompt injection
  • Data leakage
  • Agent tool misuse
  • Retrieval poisoning
  • The supporting APIs and identity

What you get

  • Findings with reproducible prompts
  • Impact per data boundary
  • Guardrail recommendations
  • A retest of fixed findings

How it runs

  • Duration: Most Assess engagements run two to four weeks from kickoff to readout.
  • What we need from you: A named decision owner, access to the assistant and the tools it can call, and the data boundaries that matter to you.

Standards

  • OWASP LLM Top 10
  • MITRE ATLAS
  • NIST AI RMF

Frequently Asked Questions

What counts as an AI system here?
The model, the agents built on it, the retrieval pipeline that feeds it and the tools it can call, together with the supporting APIs and identity.
How are findings reported?
Each one ships with the prompt or input that triggered it, so you can reproduce it, and with the impact stated per data boundary.
Does this give us AI governance evidence?
Yes. We test against the OWASP LLM Top 10, MITRE ATLAS and NIST AI RMF, and the findings and guardrail recommendations serve as that evidence.

How We Work

Eight steps from the first call to the closure letter. Every engagement follows them.

  1. Scoping call
  2. Rules of engagement
  3. Threat model
  4. Execution
  5. Analysis
  6. Readout
  7. Remediation support
  8. Retest
Read the full lifecycle

Ready to scope an engagement?

Tell us the asset and the decision the test should inform.

Scope a call