AI Security Hub • Reviewed July 2026

Use AI with confidence without surrendering control.

Practical guidance for organisations, leaders, technical teams and people navigating AI-enabled deception, sensitive data, prompt injection, connected agents and operational risk.

Explore the defence loop Open practical guides

HD secure adoption model

Six controls around every AI use case.

Move from enthusiasm to evidence. Each stage creates a decision, an owner and a control that can be tested.

DISCOVER THE REAL ENVIRONMENT

Know where AI is already working.

Inventory approved and unapproved tools, individual accounts, plugins, APIs, agents and connected information sources. Record the business purpose and accountable owner.

Map tools, users, subscriptions and integrations
Identify data sources and available actions
Assign an owner and review date

AI risk surface

Threats change. Defensive principles remain.

Use the filters to explore human, data, application, agent and supply-chain risks. Each card explains the failure mode and the first defensive move.

8 topics visible

PEOPLE / 01High believability

AI-enabled phishing and impersonation

Generated text, cloned voices and synthetic video can reproduce tone, context and authority.

First controlVerify consequential requests through an independently trusted channel.

Open verification guide

DATA / 02Disclosure

Sensitive information in prompts and connectors

Uploads, chat history, retrieval stores and connected workspaces can reveal more than intended.

First controlClassify, minimise and approve data before it enters an AI service.

Open data guide

APP / 03Instruction conflict

Direct and indirect prompt injection

Hostile instructions in webpages, documents or messages can redirect model behaviour and tool use.

First controlKeep authorisation outside the model and treat retrieved content as untrusted.

Open application guide

AGENT / 04Excessive agency

Autonomy beyond control

An agent that can email, alter files or call systems can turn a wrong decision into a real-world event.

First controlUse minimum permissions, bounded actions and approval at the point of consequence.

Open agentic AI guide

APP / 05Unsafe execution

Improper output handling

Model output may contain unsafe code, commands, links or structured data that downstream systems trust.

First controlValidate and encode output before display, storage, execution or tool use.

Review OWASP GenAI risks

SUPPLY / 06Dependency risk

Models, datasets, plugins and components

Third-party models, libraries and data pipelines can change, become compromised or inherit unknown risk.

First controlMaintain component provenance, version control, assurance and an exit path.

Open supplier guide

PEOPLE / 07Overreliance

Confident output, weak evidence

Fluent answers can be incomplete, fabricated or unsuitable for the actual context.

First controlRequire source checks and qualified human review for consequential decisions.

Browse verification guides

OPS / 08Unknown exposure

Shadow AI and unmanaged accounts

Useful tools may enter the organisation without known owners, contracts, controls or offboarding.

First controlDiscover use transparently and offer a safe approved pathway.

Open shadow AI guide

Secure adoption baseline

Five control planes. One accountable use case.

A safe AI deployment connects governance to operational evidence. The following questions create a practical starting baseline.

Use the supplier questions

01 People and accountability

Who owns the use case? Who approves consequential output? Can staff pause an unusual request without fear?

02 Data and privacy

What enters the system, where is it retained, who can retrieve it and is it used for model improvement?

03 Identity and access

Are SSO, MFA, minimum privilege, separate service identities and complete offboarding in place?

04 Technology and supply chain

Which models, plugins, APIs and datasets are used? How are changes, vulnerabilities and provenance managed?

05 Monitoring and response

Can you see prompts, retrieval, tool actions, approvals and outcomes—and can you contain the system safely?

Framework alignment

Translate principles into operating language.

The hub uses complementary models rather than claiming certification: NIST AI RMF for AI risk, ASD guidance for secure AI use, and operational cyber functions for defence.

NIST AI RMF

Govern → Map → Measure → Manage

Establish accountability, understand the use context, evaluate risk and maintain treatment throughout the lifecycle.

Official NIST resource

ASD’s ACSC

Secure adoption and AI data protection

Use Australian guidance for engaging with, deploying and protecting the data behind AI systems.

Official cyber.gov.au guidance

Operational defence

Govern → Identify → Protect → Detect → Respond → Recover

Connect AI security to the same ownership, evidence, monitoring and response disciplines used across cyber operations.

AI in cyber defence guidance

Editorial review: 15 July 2026. This material is general defensive guidance, not legal advice, a formal audit or a guarantee of security.

Make AI risk reviewable

Start with the use case that matters most.

Define the data, connected systems, decision impact and accountable owner. The right controls become clearer once the real workflow is visible.

Explore secure AI services Contact Henry