Engaging with artificial intelligence

Voluntary guidance. This is joint guidance for organisations that use AI, not legislation, and it does not bind a private law firm unless a contract or a lawful direction makes it mandatory.

Joint guidance on using AI systems securely, including systems a third party hosts. It is about secure use, not about building a secure model.

MaintainedLast reviewed 10 October 2026

Key facts

Issued by: ASD’s Australian Cyber Security Centre, with international partners
Sources listed: 3

This is a reading aid for an Australian law firm. It is not legal advice, not a certification, and not a finding that a tool or a firm complies. The sentences paraphrase the sources listed, checked on the date shown. Where a standard is sold by ISO or the AICPA, the page uses the publisher’s public description and does not reproduce the standard.

What it is

ASD’s ACSC published Engaging with artificial intelligence on 24 January 2024, with partners including CISA, the FBI, the NSA, the UK NCSC, and the Canadian, New Zealand and other national cyber centres. The publication’s purpose is to help organisations use AI systems securely. It summarises threats and prompts organisations to consider steps for managing risk. It covers self-hosted and third-party hosted systems.

The authoring agencies say the guidance is about using AI systems securely, not about developing them. Developers are pointed to the separate Guidelines for secure AI system development. ASD’s news item of the same day says the steps include applying the advice alongside the Essential Eight, knowing the system’s constraints and training staff, understanding privacy and data-protection effects, considering whether services are secure-by-design, and having suitably qualified staff.

Who it applies to

The publication is written for organisations that use, or are considering, AI systems. It is guidance. It is not an Australian statute. A law firm that uses a vendor’s chatbot, a research assistant, or a model hosted by the firm is inside the audience the paper describes. A firm that only reads about AI, and does not use a system, is not being told to adopt one.

What it asks of a law firm using AI

Start from the cyber security framework that applies in your jurisdiction. For Australia, the paper points readers to the Essential Eight and says an AI system benefits from the same mitigations used on other systems.
Consider how the system collects, processes and stores data, and how that affects privacy and data-protection obligations. The paper says AI systems are often hosted in the cloud and may send data between regions, and that the system should be able to meet data residency or sovereignty obligations.
If a third party hosts the system, understand whether the organisation’s inputs will retrain the model, and consider a private version if one is available. Also understand how data will be managed when the commercial agreement ends. The paper says that is typically in the privacy policy or terms of service.
Be cautious about what people type into a generative system. The paper says information given to these systems may be incorporated into training data and could inform outputs to people outside the organisation.
Require phishing-resistant multi-factor authentication for access to the AI system and to repositories that hold training data. Limit privileged access on a need-to-know and least-privilege basis.
Train staff on what must not be entered, including personally identifiable information and the organisation’s intellectual property, and on how far outputs can be relied on. The paper says generative systems are known to hallucinate, and it uses a 2023 New York court filing as the example.
Plan for failure. If a third party runs the system, know the vendor’s availability commitments and who does what in an incident. Log inputs and outputs. Revisit the questions as the technology changes. The paper says there are limited regulations to ensure AI systems are secure, so the organisation has to consider the risks itself.

How it maps to this framework

The rows are this site’s questions. Where the cell says “Not set here”, the instrument does not set that requirement. The gap is still recorded, so it is not treated as answered.

Mapping Engaging with AI onto this site. Checked 2026-10-10.
This site’s question	What the instrument says that bears on it
Efficiency Whether the tool saves total human work at the quality the task needs.	ASD’s announcement says that, used securely, AI has the potential to improve efficiency and effectiveness. The paper does not measure whether a particular legal task gets faster. A saving still has to be timed on the firm’s own work.
Effectiveness Whether the output is good enough for the task. The catalogue does not score this.	The paper treats hallucinations, data poisoning and prompt injection as reasons an output can be wrong or malicious. That supports checking the output. It does not score a tool’s legal effectiveness.
Due diligence What the firm has actually read about hosting, training and assurance before it relies on the tool.	The questions in the paper are a due-diligence list: residency, retraining, exit, supply chain, limits of the system, staff competence, logging, and incident duties. This framework’s due-diligence level is the firm’s reading of those facts, not a box the paper ticks.
Confidentiality Whether client and other confidential information can go into the tool.	The paper says organisations should be cautious with sensitive data, customer personal data and intellectual property, because prompts may be used to train the system and later appear in someone else’s output. That is a security warning. The solicitor’s duty is on the confidentiality page.
Disclosure When the firm tells the client, and where required the court or another party, that AI was used.	Not set here.
Certifications ISO, SOC 2 and IRAP names recorded against the tool. A listed name is not a finding that the certificate covers the firm’s plan.	The paper does not treat an ISO or IRAP badge as a substitute for the questions it asks. It does say to seek vendors that are transparent about how they developed and tested the system, and to consider the NCSC-UK guidelines for secure AI system development.
Data residency Where the tool holds data, including whether that place is recorded as Australia.	The paper states the point directly: ensure that any AI system the organisation uses can meet its data residency or sovereignty obligations, because cloud systems may send data between regions. Record where the data sits. The paper does not name Australia as the only acceptable place.

Sources

Checked 2026-10-10. Issued by ASD’s Australian Cyber Security Centre, with international partners.

How this framework decides Software catalogue All frameworks