Build AI. Secure AI. AI security for Australian organisations.

WIT Solutions provides AI security services across Australia. We test chatbots, copilots and LLM applications the way an attacker would. We review Microsoft 365 Copilot before it surfaces the wrong files, and help you set AI governance that staff will follow.

Independent and practitioner-led. No products to resell, and the person who scopes your work does the testing.

  • Mapped to the OWASP Top 10 for LLM Applications
  • Microsoft 365 Copilot readiness reviews
  • Independent and practitioner-led
  • Australia-wide, remote or on-site

Three sides of AI security

Using AI to defend, securing the AI you already run, and building new AI safely.

AI-enhanced security

We use AI to speed up our own security work. A person reviews every result before it reaches you.

  • Faster log analysis during investigations
  • AI-assisted triage of alerts and findings
  • Quicker, clearer reporting
  • A human in the loop on every decision

Securing your AI systems

We attack your chatbots, copilots and LLM apps the way a real adversary would. Then we help you close the gaps.

  • LLM and chatbot penetration testing
  • Prompt injection and jailbreak testing
  • Microsoft 365 Copilot oversharing reviews
  • Shadow AI discovery and AI governance

Secure AI development

Building with AI? We help you design RAG apps, agents and automations with security built in from day one.

  • Security architecture review for RAG and agents
  • Least-privilege access to tools and data
  • Safe handling of inputs, outputs and secrets
  • Security testing before go-live

Why AI security is different from a traditional penetration test

A traditional penetration test looks for flaws in code and configuration. We still test those. But AI applications add risks that scanners and checklists miss.

Instructions and data share one channel

An LLM reads its instructions, the user’s message and retrieved documents as one stream of text. It can’t reliably tell commands from content, so anyone who can put text in front of it may steer it.

Attacks can arrive indirectly

A hidden instruction can sit in an email, PDF, web page or support ticket. When the AI reads it, the attack runs. The attacker never needs to touch your chatbot.

Outputs aren’t predictable

LLMs are non-deterministic: the same prompt can give different answers each run. One clean test proves little, so we repeat and vary attacks.

AI now has access and agency

Copilots can search everything a user can see. Agents can send email, call APIs and change records. The more an AI can do, the more damage one injection can cause.

So we test AI the way we test everything else: as attackers, with evidence. Our offensive security practice can test the APIs and authentication around it in the same engagement. Findings feed into your governance, risk and compliance program and our defensive security work.

The OWASP Top 10 for LLM Applications, in plain English

We map findings to the OWASP Top 10 for LLM Applications (2025 edition), a widely used list of the main LLM security risks.

  1. LLM01

    Prompt Injection

    Crafted input, typed in or hidden in content, makes the model ignore its instructions.

  2. LLM02

    Sensitive Information Disclosure

    The model reveals personal data, secrets or confidential business information in its answers.

  3. LLM03

    Supply Chain

    Weaknesses in the third-party models, datasets, plugins and libraries you build on.

  4. LLM04

    Data and Model Poisoning

    Tampered training or fine-tuning data plants bias, backdoors or harmful behaviour.

  5. LLM05

    Improper Output Handling

    Model output is passed to browsers, databases or other systems without checks, opening the door to classic attacks.

  6. LLM06

    Excessive Agency

    The AI has more permissions, tools or autonomy than its job needs, so one mistake or attack does real damage.

  7. LLM07

    System Prompt Leakage

    Hidden instructions are extracted, exposing logic, filters or secrets that never belonged in a prompt.

  8. LLM08

    Vector and Embedding Weaknesses

    Flaws in RAG retrieval let users pull documents they shouldn’t see, or let attackers plant content.

  9. LLM09

    Misinformation

    The model states false information with confidence, and people or systems act on it.

  10. LLM10

    Unbounded Consumption

    Uncontrolled use runs up costs, degrades the service or helps someone copy the model.

We also map attack techniques to MITRE ATLAS, a public knowledge base of adversary tactics and techniques against AI systems. It gives your teams a shared language for fixes and detection.

Our AI security services

From LLM security testing to Microsoft 365 Copilot reviews and AI governance. Use one, or combine them.

LLM penetration testing and AI red teaming

Hands-on testing of the chatbots, copilots, RAG apps and agents you build or deploy. We try to make them leak data, break their rules or misuse their access.

Best for: Customer-facing chatbots and internal AI apps, before launch or after major changes.

Show what's included
  • Direct and indirect prompt injection
  • Jailbreaks and system prompt extraction
  • Data leakage across users, documents and tenants
  • Abuse of connected tools, plugins and APIs
  • Authorisation bypass in the app around the model

Microsoft 365 Copilot readiness and oversharing review

Copilot can surface anything a user already has access to. We find the overshared SharePoint, OneDrive and Teams content it would expose, and help you fix it before rollout.

Best for: Organisations planning, piloting or expanding Microsoft 365 Copilot.

Show what's included
  • Permissions and sharing-link review
  • Sensitivity labels and Microsoft Purview DLP
  • Options to limit what Copilot can search
  • A staged, measured rollout plan

Shadow AI discovery and AI acceptable use policy

Find out which AI tools staff actually use and what data goes into them. Then put a practical policy in place.

Best for: Any organisation without a clear view of how staff use AI.

Show what's included
  • Discovery using your existing network, identity and cloud app data
  • Short, non-punitive staff survey
  • Risk ranking of each tool by the data it receives
  • AI acceptable use policy and approved-tools list

AI governance and risk assessment

A practical AI governance framework sized to your organisation, with controls mapped to recognised frameworks.

Best for: Boards, risk owners and regulated organisations that need to show AI is under control.

Show what's included
  • AI inventory and risk register
  • Gap assessment against ISO/IEC 42001:2023
  • Alignment with the NIST AI Risk Management Framework
  • Australian Government guidance on safe and responsible AI
  • Privacy review against the Australian Privacy Principles

Secure AI development and architecture review

Security design review for RAG apps, agents and AI automation, before you build or once they’re live.

Best for: Product and development teams building with LLMs.

Show what's included
  • Least-privilege access to tools and data
  • Safe input and output handling
  • Secrets management, logging and monitoring
  • Human approval for high-impact actions
  • Evaluation and security testing before release

AI agent and integration security

Agents can read email, call APIs and update records. We test what an agent can be tricked into doing, then tighten its permissions.

Best for: Teams connecting AI to business systems through agents, plugins or MCP.

Show what's included
  • Excessive agency and permission scoping
  • Review of Model Context Protocol (MCP) servers and connectors
  • Identity, tokens and delegated access
  • Approval steps for risky actions

AI-enhanced security operations

We use AI to speed up our own analysis, with a person checking every result. Faster answers, not unchecked automation.

Best for: Clients of our defensive security and incident response work.

Show what's included
  • Faster log analysis
  • Alert and findings triage
  • Quicker, clearer reports
  • Human review of every result

Microsoft 365 Copilot security: fix oversharing first

Copilot doesn’t change who can open a file. It changes how easy that file is to find.

Microsoft 365 Copilot respects your existing permissions. It answers questions using anything the user can already reach in SharePoint, OneDrive, Teams and email.

The catch is that permissions build up over years. Files shared too widely used to be hard to find. With Copilot, they are one question away.

So Copilot readiness is mostly a permissions and data governance job. We review and fix, then test with realistic prompts from an ordinary user’s account to show what Copilot would reveal.

Some controls depend on your Microsoft 365 licensing. We’ll tell you which apply.

Ask about a Copilot readiness review

What our Copilot review covers

  • Sites, Teams and libraries open to “Everyone except external users”
  • “Anyone” links and organisation-wide sharing links that are still active
  • Sensitive files without sensitivity labels, and how encrypted labels limit what Copilot can use
  • Microsoft Purview DLP policies, including keeping labelled content out of Copilot
  • Restricted SharePoint Search and other ways to limit Copilot’s reach during a pilot
  • A staged rollout plan with a small pilot group and clear go/no-go checks

How an AI security assessment works

From first call to retest. You know the scope and price before testing starts.

  1. Scope the system and its data

    We map what your AI can see and do: models, prompts, data sources, tools and users. You get a fixed-price quote.

  2. Threat model

    We work out who might attack it, what they would want and which paths matter most, mapped to OWASP and MITRE ATLAS.

  3. Test

    Manual adversarial testing by an experienced tester, backed by automated probing to cover many prompt variations.

  4. Report

    Every finding includes reproducible prompts, evidence, business impact and a clear fix. Leaders get a plain-English summary.

  5. Fix and harden

    We work with your developers or IT team on fixes: permission changes, guardrails, output handling and monitoring.

  6. Retest

    We retest the fixes and update the report, giving you evidence for auditors and boards.

AI systems tested by offensive security practitioners

Plenty of AI risk reviews happen on paper. Ours start with an attempt to break the system.

Tested, not just reviewed on paper

Your AI systems are tested hands-on by an offensive security practitioner. We show what an attacker could actually do, with evidence.

Independent advice

We don’t resell AI products, licences or platforms, and we’re not an MSP. Our advice isn’t shaped by what we would earn from it.

One practitioner, start to finish

The person who scopes your work is the person who does the technical work. Nothing gets lost in a hand-off.

Australia-wide

Delivered remotely or on-site anywhere in Australia, with Australian privacy, regulatory and government security requirements in mind.

Credentials held

Hands-on offensive and defensive certifications earned through practical exams.

  • OSCP
  • OSWE
  • OSED
  • OSEP
  • OSCE3
  • CRTO
  • CRTL
  • OSDA

AI security use cases by industry

We work with Australian SMBs, councils, regulated industries, government suppliers and enterprises.

Financial services

Customer chatbots and internal assistants often sit close to account data. We test for leakage and authorisation bypass, and help APRA-regulated entities bring AI into their CPS 234 and CPS 230 programs.

Healthcare

Health information is sensitive information under the Privacy Act, with stricter rules. We check what patient data reaches AI scribes, chatbots and staff tools, and where it ends up.

Local government

A council chatbot must not leak records or confidently give residents wrong answers. We test public-facing AI, review Copilot rollouts against records and privacy obligations, and help write AI policy.

Legal and professional services

Confidentiality and legal professional privilege still apply when staff use AI. We find where client data goes, test document AI and Copilot permissions across matters, and set a clear policy.

Government suppliers

Government buyers may ask how you use AI with their data. We help you answer with a policy, documented controls and test evidence.

Enterprise

Copilots, internal RAG search, agents and vendor AI features, often all at once. We help you set governance that scales, and test the highest-risk systems first.

AI security FAQs

Straight answers to common questions about AI security, Microsoft 365 Copilot and AI governance in Australia.

What is prompt injection?

Prompt injection is text that makes an AI model ignore its instructions. Direct injection is typed in by a user. Indirect injection hides in content the AI reads, such as a document, email or web page. It is first on the OWASP Top 10 for LLM Applications. There is no single fix: limit what the AI can see and do, then test it.

Is Microsoft 365 Copilot safe to roll out?

Copilot respects your existing Microsoft 365 permissions, so it is only as safe as they are. Files open to the whole organisation, old ‘anyone’ links and broad Teams membership become one question away. Review permissions, label sensitive content and roll out in stages, and the risk becomes far easier to manage.

How do we stop staff putting sensitive data into ChatGPT?

Blocking AI sites alone often pushes staff to personal phones and tools you can’t see. First, find out which AI tools are in use and why. Then offer an approved tool, a short policy on what data can go in, and controls such as DLP and web filtering for the riskiest data.

What is AI red teaming?

AI red teaming is adversarial testing of an AI system. The tester tries to make it leak data, break its rules or misuse its connected tools. We focus on security outcomes, such as data exposure and unauthorised actions, not just offensive content. Testing is hands-on, backed by automated tools for scale.

How do you test an AI chatbot for security?

We start with what the chatbot can reach: data, APIs, tools and user roles. Then we try direct and indirect prompt injection, jailbreaks, system prompt extraction, cross-user data leakage and abuse of any actions it can take. Every finding includes the exact prompts to reproduce it.

Do we need an AI acceptable use policy?

If staff use AI tools at work, or might, yes. A good policy is short. It lists approved tools, says what data must never go in, and tells people who to ask. Back it with training and technical controls, and review it as tools change.

What is ISO 42001 and do we need it?

ISO/IEC 42001:2023 is the international standard for an AI management system, much as ISO/IEC 27001 is for information security. Certification is voluntary, and usually only worth pursuing if customers or contracts ask for it. We run gap assessments and help build the controls; certification itself comes from an accredited certification body.

Does the Privacy Act apply to AI tools?

Yes. The Privacy Act 1988 and the Australian Privacy Principles apply to personal information, with or without AI. Putting customer or staff data into an AI tool can be a use or disclosure, may send it overseas, and it must stay secure. The Act covers most businesses with turnover above $3 million, all health service providers and federal agencies. Councils usually fall under state or territory privacy laws. This is general information, not legal advice.

Can you test AI tools we bought rather than built?

Often, yes. We test how a vendor’s AI is set up in your environment: what data it can reach, who can use it and what it can do. Testing the vendor’s own platform usually needs their permission, so we check their terms first.

How much does an AI security assessment cost?

It depends on scope: how many AI systems, how complex they are, what they connect to, and whether you need testing, governance or both. A single chatbot test is a smaller job than a Copilot review across a large tenant. We give a fixed-price quote after a free scoping call.

Find out what your AI could be tricked into doing

Book a free scoping call. We’ll map your AI systems and data, then give you a fixed-price quote.