AI Security
AI and LLM Application Security Testing
Your chatbot can be persuaded to break its own rules
A crafted message can make an AI assistant reveal its instructions, leak customer data it can reach, or take actions it should not, such as sending messages or changing records. Ordinary scanners do not test for any of this.
Prompt injection
Direct and indirect attempts to override the assistant's instructions, including through documents and web pages it reads.
Data leakage
Whether the AI can be made to reveal other customers' data, internal documents or its system instructions.
Unsafe actions
If the AI can call tools or APIs, whether it can be tricked into actions it should not take.
Abuse and cost
Whether attackers can misuse it for spam, harmful content or to run up your AI bill.
How we work
01
Scope
We agree which AI features, data and tools are in scope, and you sign a written authorisation.
02
Map
We map what the AI can read, remember and do, since that decides the real risk.
03
Test
Injection, leakage and misuse attempts, run carefully against the agreed environment.
04
Report
Findings ranked by business risk with clear fixes, then a call to agree priorities.
What you get
- Written authorisation and an agreed scope before any testing
- Prompt injection testing, direct and through content the AI reads
- Data leakage and system prompt exposure checks
- Review of tools, permissions and actions the AI can take
- A written report with findings ranked by business risk
- A walk through call and a practical fix plan
AI and LLM application security test
Quoted
Priced after scoping, based on the AI features, the data and the tools they can reach. Work starts once the commitment fee in the quote is paid.
Before you approve it
The engagement, in facts.
- Typical timeline
- Agreed at scoping
- Report
- Within 3 working days of testing ending, findings ranked by risk with a fix for each
- Retest
- Free retest of fixed findings within 30 days
- Standards
- OWASP, PTES, NIST SP 800-115 and CVSS
- Confidentiality
- We sign your NDA before scoping, on request
Questions, answered
What is prompt injection?
A message, or content the AI reads, that tricks it into ignoring its instructions. It is the most common way AI features are attacked, and a better prompt alone cannot fully fix it.
Which standard do you follow?
The OWASP Top 10 for Large Language Model Applications guides the testing, alongside the usual web and API checks for the systems around the AI.
Do you test on our live system?
Testing is agreed in the scope. A staging copy is preferred; if it must be live, risky actions are limited and agreed with you first.
We use an AI provider such as OpenAI or Anthropic. Is that in scope?
The provider's own systems are not tested. What is tested is your integration: your instructions, the data and tools you connect, and how your app handles the AI's answers.