AI era service

LLM and AI agent security testing

Your AI feature follows instructions. Whose?

Chatbots, copilots, retrieval pipelines and agents with tools, tested for prompt injection, data leaks and over-powered tools.

  • Prompt injection
  • Tool permissions
  • RAG and memory
  • MCP servers
Book a scoping callWhat we look at

Scope and deliverables

What we look at

  • Direct and indirect prompt injection
  • What each tool can reach, and on whose authority
  • Leaks between users through context, memory or retrieval
  • System prompt and secret exposure

What you get

Attack transcripts for every issue and concrete guardrails for your stack.

Typical length

1 to 2 weeks

Methods aligned with

OWASP Top 10 for LLM Apps, CVSS 3.1

Why it matters

Chatbots, copilots and agents read text from places you do not control: support tickets, uploaded documents, web pages and email. If that text carries instructions, the model may follow them and use its tools to act on them.

We test what your AI features can be talked into, what each tool can reach, and whether one user can pull another user’s data out through context, memory or retrieval.

Where AI features break

  • 01

    Indirect prompt injection

    Instructions hidden in a ticket, document or web page that the model reads and then follows.

  • 02

    Over-powered tools

    A tool that can read or change more than the person using it is allowed to.

  • 03

    Leaks between users

    One user’s data surfacing in another user’s answer through shared context, memory or retrieval.

  • 04

    System prompt and secret exposure

    Hidden instructions, keys or internal details the model can be talked into repeating.

How we test it

Hackbots for breadth. Hackers for depth. Every lead is proven or discarded by a researcher before it reaches your report.

Our hackbots

  • Hit every AI feature with thousands of injection variants
  • Call each tool the model can reach, as each role
  • Re-run the full set when the model, prompt or tools change

Our researchers

  • Trace what each tool can reach, and on whose authority
  • Chain an injection into real data access, not just an odd reply
  • Write concrete guardrails for your stack

How an engagement runs

Five steps from the first call to the retest. You work directly with the researchers testing your systems at every one of them.

  1. Scope

    1 to 3 days

    We agree on assets, test accounts and rules of engagement, then send a fixed price.

  2. Recon

    First days of testing

    Hackbots map every host, route and role. A researcher builds the threat model.

  3. Exploit

    Most of the window

    Hackbots run the matrix at scale while researchers test by hand and take each lead to the end.

  4. Report and fix

    Within 5 working days

    Verified issues only, each with a CVSS 3.1 vector, replayable proof and a patch.

  5. Retest

    Free, within 90 days

    We confirm each fix and issue an updated report and attestation letter.

All 9 services

Tell us what you run. We’ll send a fixed quote.

No obligation and no sales sequence. A senior researcher reads every request and replies within one working day.

Book a scoping call[email protected]