AI Red Team Testing
We try to break your AI systems, with your permission, before someone else does.
Red Team Testing
Find the Weak Spots Before an Attacker Does
A chatbot on your website, an assistant connected to your inbox or files, an agent that can take actions in your systems: each one takes instructions in plain language, and anyone who can type to it can try to give it different instructions. Red team testing means we play the attacker, find out what your AI can be talked into, and show you how to stop it.
Every test runs under signed rules of engagement: scope, timing, safety and data handling.
The Details
What's Included
← All AI Security & Governance services
What we test for
Testing follows the OWASP Top 10 for LLM Applications, and findings are described using MITRE ATLAS so your team and your vendors can act on them.
- Prompt injection. Direct attempts to override the system's instructions, and indirect ones hidden in web pages, emails or documents the AI reads.
- Data leakage. Getting the system to reveal customer records, personal information, internal documents, or its own hidden instructions.
- Excessive agency. Tricking an assistant or agent into sending, changing or deleting things it should not, using the access it was given.
- Jailbreaks and unsafe output. Getting the system to produce content that would embarrass you or create liability.
- Insecure output handling. AI output that is passed straight into other systems, web pages or automations without being checked.
- Access and isolation. Whether one customer or user can reach another's data through the AI.
Rules of engagement
Security testing is only legitimate when it is authorized. Every engagement runs under a signed agreement that sets out:
- Written authorization from someone with authority over every system in scope, including confirmation that testing is allowed under the terms of any third-party AI or software vendor involved.
- Scope. Exactly which systems, accounts and environments may be tested, and what is off limits.
- Timing. Agreed testing windows, and a named contact on your side we can reach immediately.
- Safety. Testing in a non-production environment wherever one exists, test accounts instead of real customers, and a stop condition if anything risks harming live operations.
- Data handling. Any real data we encounter is touched only as far as needed to prove a finding, never copied beyond the evidence, kept encrypted, and destroyed at the end of the engagement, with written confirmation.
- Confidentiality. Findings are shared only with the people you name, through our secured client portal.
Who does the testing
Larry tests the AI systems: the chatbots, assistants, agents and AI-powered automations. When an engagement also needs traditional network or web application penetration testing, that work goes to specialist partners, who are named in your proposal before you sign, and who work under the same rules.
What you get
- A findings report ranked by severity, with the evidence for each finding and how to reproduce it.
- Specific fixes for each finding, written for whoever maintains the system.
- An executive summary in plain English.
- A retest after you have made the fixes, to confirm they hold.
Get Started
Start With a Scoping Call
Thirty minutes, by video or on-site. We will talk through how your business uses AI today and where customer data goes, and you will leave with a written scope and quote. Nothing is tested and no data is shared until a written agreement is signed.
