AI Penetration Testing

Security testing for LLM applications, AI agents and MCP servers.

AI penetration testing is manual security testing of the AI systems your business has connected to its data and its tools: LLM applications, AI agents, and Model Context Protocol (MCP) servers. It looks for the weaknesses ordinary application testing does not cover, chiefly prompt injection, excessive agency, and data leaking between tenants. Siege Cyber is a CREST accredited Australian firm based in Brisbane, and we test these systems the way an attacker would.

Why AI systems need their own penetration test

Single sign-on, conditional access and IP restrictions control who can log in. They do nothing about prompt injection, because the attack never arrives at the login screen. It arrives inside content your AI reads: a helpdesk email, a support ticket, a shared document, a web form, or a page your agent browses.

One malicious ticket can carry hidden instructions. If your agent reads it and holds tools such as your RMM, your ticketing system or Microsoft 365, it can act on the attacker’s behalf. That can mean leaking customer data, changing a configuration, or pushing software across your estate. The agent does exactly what it was built to do, which is why your existing controls do not stop it.

What we test

Testing is mapped to the OWASP Top 10 for LLM Applications 2025, published by the OWASP GenAI Security Project, and to Securing Agentic Applications V2 where your system uses agents. We combine manual testing by experienced penetration testers with automated tooling.

OWASP risk What we test for
LLM01:2025 Prompt Injection Direct and indirect injection across every channel your AI reads: email, tickets, documents, web forms and pages an agent browses.
LLM02:2025 Sensitive Information Disclosure Whether the model will reveal customer data, credentials or internal detail it holds or can reach.
LLM03:2025 Supply Chain The models, plugins, connectors and third-party services your AI depends on, and what they can reach.
LLM04:2025 Data and Model Poisoning Whether untrusted content can reach training, fine-tuning or retrieval data and change later behaviour.
LLM05:2025 Improper Output Handling What happens downstream when model output is passed to a browser, a shell, a database or another system.
LLM06:2025 Excessive Agency What actions an injected prompt can actually trigger, and whether risky ones require a human to approve.
LLM07:2025 System Prompt Leakage Whether your system prompt, rules or guardrails can be extracted and then worked around.
LLM08:2025 Vector and Embedding Weaknesses Retrieval and vector store security, including whether one tenant can reach another tenant’s documents.
LLM09:2025 Misinformation Where confident but wrong output could cause real harm, and whether anything checks it.
LLM10:2025 Unbounded Consumption Whether an attacker can drive cost or denial of service through your AI endpoints.

Our AI testing services

Service What it covers
AI and LLM application penetration testing Manual testing of your AI-powered application end to end, mapped to the OWASP Top 10 for LLM Applications 2025.
MCP server and connector review Authentication, access control, token handling, secrets storage and permission scope on the MCP servers and connectors your AI uses.
AI security architecture review and threat modelling A review of your agents, connectors, tool permissions and trust boundaries before or alongside testing, to find design-level problems that testing alone will not surface.

How this differs from a web application penetration test

A web application penetration test asks what a user can reach that they should not. An AI penetration test asks what your application can be persuaded to do on someone else’s instructions. The two overlap, and most AI applications need both: the AI layer sits on top of an ordinary web application that still has ordinary flaws.

If your AI feature is new and the application underneath it has never been tested, we will usually say so and recommend testing the application first. It is the cheaper problem to find.

What you receive

  • A report written in plain business terms, showing exactly which input produced which action.
  • A severity rating per finding, with practical remediation your team can act on.
  • A walkthrough of the findings with the tester who did the work.
  • A retest of the fixes, so you can show the issues are actually closed.

Who this is for

Software and SaaS companies building AI features into their product. Managed service providers running AI agents that hold access to client environments. And any Australian organisation connecting AI tools to systems that hold customer or personal information.

Common questions

What is AI penetration testing?

AI penetration testing is security testing of an AI system: the LLM application, the agent, and the connectors and MCP servers it uses. Rather than only testing who can log in, it tests what the system can be made to do by content it reads, which is where the risk in AI applications actually sits.

What is prompt injection?

Prompt injection is when an attacker hides instructions inside content an AI system processes, causing it to ignore its intended rules or take unintended actions. Indirect prompt injection is the same attack delivered through a document, email or web page the AI reads rather than typed by the user.

What is an MCP server, and why does it need testing?

The Model Context Protocol lets AI assistants connect to tools and data sources. An MCP server is what exposes those tools, which makes its authentication, token handling and permission scope a security boundary. If the scope is too wide, a successful prompt injection reaches further.

Do SSO and conditional access protect against prompt injection?

No. They control who can sign in. Prompt injection arrives through content the AI reads after a legitimate user is already signed in, so the login controls are never tested by the attack.

How often should AI systems be tested?

At least annually, and after any significant change to the model, the prompts, the tools an agent can call, or the data it can reach. For agents that can take actions in your environment, more frequent testing is sensible because the blast radius is larger.

Does this cover ISO 42001?

An AI penetration test produces evidence that supports an AI management system, but it is not a certification activity. If you are working toward ISO 42001 certification in Australia, our compliance team handles that separately.

Connecting AI to your business systems?

Book a short call with our team to talk through what an AI penetration test would cover for your environment.

Book a meeting