Threat model
How your AI system could be attacked, by whom, and what is at stake.
Service 06
AI security testing
Adversarial testing of your chatbots, copilots and AI agents for prompt injection, data leakage and excessive permissions, before attackers or customers find the gaps.
Why it matters
AI is your newest attack surface.
Once an AI system can read your data and call your tools, a single malicious instruction hidden in an email, document or web page can make it act against you. Traditional penetration tests rarely cover this.
- Find leaks of customer data and system instructions
- Prove agents cannot be hijacked into harmful actions
- Give your board evidence that AI risk is managed
What we test
Mapped to the OWASP Top 10 for LLMs.
Prompt injection
Direct and indirect attacks through chat, documents, email and web content.
Data leakage
Exposure of personal information, secrets and system instructions.
Excessive agency
Tool permissions, approvals and what an agent can really do on its own.
Retrieval and vector stores
Poisoned content, cross-tenant access and permission bypass in search.
Remediation and retest
Prioritised fixes, help implementing them, and a retest to confirm.
Engagement
AI Security Assessment
Two to three weeks. Threat model, hands-on testing, an executive summary for the board and a technical report your engineers can act on.
from A$12,000
What is prompt injection?
An attack where instructions hidden in user input, documents, web pages or emails steer an AI system into doing something it should not, such as revealing data or misusing a connected tool. Read our guide for leaders.
Is testing safe for our production systems?
We agree scope, rules of engagement and test environments in writing first. Most testing runs against staging, and anything touching production is coordinated with your team.
Do you help fix what you find?
Yes. Every finding comes with a practical fix, and we can implement the remediation and retest to confirm it worked.
How often should we test?
Before launch, after significant changes to models, prompts, tools or data sources, and at least annually for systems in production.
Related