AI Security & Reliability
LLM & AI Agent Security Testing
Find out how your chatbot, copilot, agent or RAG feature could be misused, before someone else does.
- Senior engineers only
- NDA on request
- Reply within 1 business day
Fit
Who it's for
A good fit if one of these sounds like you.
Teams shipping chatbots, copilots or AI assistants
Find out what users can make your assistant say, reveal or do.
Products with AI agents that call tools or APIs
Confirm an agent can't be steered into actions it shouldn't take.
Companies adding RAG features over internal or customer data
Make sure retrieval only returns what each user is allowed to see.
Scope
What we do
Prompt-injection testing
Direct and indirect injection through user input, documents, web pages and tool results.
Data-leakage tests
Attempts to extract system prompts, other users' data and retrieved documents.
Tool and function-call abuse
Whether the model can be steered into calling tools it shouldn't, with inputs it shouldn't.
Agent permission review
Least-privilege review of what each agent can read, write and trigger.
Output handling checks
How model output is rendered, stored and passed to other systems.
Retest after fixes
We verify each fix against the original reproduction.
Deliverables
What you get
Concrete deliverables your team keeps and can build on.
- Report with reproductions for every finding
- Severity ratings and recommended fixes
- Agent and tool permission map
- Retest after fixes
Process
How it works
- 1
Scope
We map your AI features, models, tools and data sources, and agree what's in scope.
- 2
Threat model
We identify who could attack each feature, how, and what they could reach.
- 3
Test
Structured and exploratory testing against real attack patterns.
- 4
Report & retest
Findings with reproductions and fixes, then a retest once you've applied them.
FAQ
Frequently asked questions
Which models and providers do you test?
Features built on the major LLM providers and on open-source models. What matters most is how the model is wired into your product, which is what we test.
Do you test in production?
We prefer a staging environment with realistic data. If testing in production is needed, we agree limits in writing first.
Can prompt injection be fully prevented?
Not with prompts alone. We focus on limiting what a successful injection can do: permissions, output handling and human approval for sensitive actions.
How often should we test?
Whenever you add a new tool, data source or model change. With a retainer, testing can run on every release.
Insights
Related insights
AI Security
LLM Output Handling: How to Render AI Responses Safely
LLM output handling in practice: render responses as text, block markdown image leaks, validate structured output and keep it out of SQL and shells.
7 min read
AI Security
RAG Access Control: How to Stop Document Leaks
RAG access control in practice: enforce document permissions inside the search query, sync revocations, isolate tenants and test for leaks before launch.
7 min read
AI Security
How to Test an LLM Feature Before Launch
How to test an LLM feature before launch: what to attack, how to write test cases you can rerun, and the pass criteria to agree before go-live.
6 min read
Get in touch
Tell us what you're building.
Share a few details and a senior engineer will reply within 1 business day.
Prefer to talk? Book a call
Or email contact@geektech.com