Preparing for Advanced AI: What Households Can Reasonably Do

Concerns around AI are increasingly front of mind for many clients, who are left wondering what they actions they can realistically take to mitigate risks associated with advanced artificial intelligence.

“AI risk” covers a wide range of very different possibilities. Some are already visible: AI-assisted cyberattacks, fraud and impersonation, automated influence operations, and the growing practice of giving AI systems access to consequential real-world functions. Others are developing rapidly—increasingly autonomous cyber capability, AI-assisted biological research resulting in engineered pathogens and advanced bioweapons, autonomous military systems, and agentic software operating critical systems. At the far end lie genuine loss-of-control scenarios, where both probability and consequences remain uncertain and private preparedness becomes increasingly speculative.

For people planning homes, properties or continuity infrastructure with a 10–30-year life, the practical task is to distinguish present risks from plausible future developments, and those again from genuinely unknowable scenarios. The focus should remain on the pathways through which AI-related events could actually reach the household. Many quite different AI risks ultimately converge on similar practical consequences: power disruption, communications failure, payment and logistics problems, pressure on healthcare, loss of trusted information, reduced emergency response, and possible deterioration in public order.

Looking Beyond Today’s Models

Hardware, infrastructure and other preparedness measures implemented today may be expected to still be effective in 2040 or 2050. Current AI capability is therefore a poor upper bound for the environment those measures may face. UK AI Security Institute testing since 2023 shows rapid improvement in cyber capability, biology and chemistry, and autonomous task completion. On AISI cyber evaluations, the length of tasks frontier models can complete without assistance has been increasing rapidly—earlier estimates around eight months, with more recent assessments closer to four to five months on some metrics—though performance remains inconsistent on long, realistic attack chains. AISI has also measured substantial gains in laboratory troubleshooting, biological reasoning and experimental support.

The UK National Cyber Security Centre expects AI to increase the effectiveness and efficiency of cyber intrusion, raise attack volume, particularly benefit highly capable state actors, and improve vulnerability discovery, exploit development, reconnaissance and automation. Important caveats apply: benchmark progress cannot simply be extrapolated mechanically to 2031 or 2041; defensive AI will also improve; and real infrastructure is much harder to attack than benchmark environments. The sensible planning assumption is straightforward: do not put in place twenty-year preparedness measures based on the limitations of 2026 AI.

Cyberattack and Infrastructure Disruption

This is probably one of the clearer AI-related pathways to household disruption in Australia and New Zealand. Sophisticated cyber operations currently require large amounts of specialist human labour—finding targets, analysing software, discovering vulnerabilities, maintaining access, processing stolen information and adapting as defenders respond. More capable AI could automate parts of that work, allow many more systems to be examined, reduce specialist labour bottlenecks, accelerate exploit development and enable more operations to run in parallel.

State actors are particularly important because they may already possess stolen credentials, compromised supply chains, human intelligence, zero-days, persistent access inside target networks and specialist cyber teams. AI can act as an analysis and coordination layer over those existing capabilities.

Household concern is broader than “the power grid gets hacked.” Critical systems are interdependent—electricity, telecommunications, payments, water, fuel, freight and logistics, and healthcare. A coordinated or cascading disruption could affect several simultaneously. Potential household effects include grid failure, mobile and internet outages, EFTPOS and payment problems, fuel shortages, water-system issues, freight delays, refrigeration loss, and impaired healthcare and emergency response.

There is also a non-malicious pathway. As AI agents increasingly receive authority over real systems, poor design, bad instructions or software errors could propagate disruption at machine speed without an adversary. Australian Signals Directorate guidance already emphasises isolating operational technology, maintaining essential services during cyber incidents, and reducing external dependencies.

Information, Identity and Trust

AI is progressively weakening familiar indicators of identity—voice, photographs, writing style, video, caller ID and account identity. It can also combine social-media history, leaked databases, business information, family relationships, travel history, property information and existing voice or video material. This makes impersonation and manipulation increasingly personalised.

At the individual level this can appear as a fake call from a child or spouse, a fabricated business instruction, a compromised real email account combined with a cloned voice, or an apparently genuine video message requesting urgent action. Larger-scale influence operations can mix synthetic media, hacked real accounts, leaked genuine documents, fabricated documents and coordinated automated accounts.

The deeper risk is not necessarily persuading everyone to believe one false story, but making it harder for people to determine what is genuine at all. Highly personalised deepfake calls or video from a supposed family member, police officer or official can create immediate physical vulnerability—prompting people to leave secure locations, open doors, move valuables or travel to a designated place under false pretence of kidnapping, home invasion or medical emergency. At larger scale, synthetic evidence of institutional collapse or betrayal (deepfakes and fabricated documents appearing to show police, military or local authorities turning against the population or having already abandoned their posts) can trigger pre-emptive defensive mobilisation, armed vigilantism or widespread refusal to cooperate with legitimate responders. During a real or suspected outbreak, AI-generated “expert” or official-looking guidance can misdirect people to the wrong treatment centres, promote harmful self-medication, or declare certain suburbs or facilities safe when they are not. Credible-looking reports that banking or payment systems will be offline for weeks, reinforced by synthetic “insider” confirmation, can produce simultaneous mass cash withdrawals, fuel stockpiling and supermarket runs that create genuine shortages and secondary disorder even if the original claim is false. Repeated successful false alarms delivered through compromised or synthetic channels can further erode trust in warning systems, so that when a genuine high-impact event occurs the population’s response is delayed or partial and physical harm is multiplied.

Infrastructure disruption and information disruption can reinforce each other. Real outages create uncertainty; false information becomes harder to check; authorities may struggle to establish a shared picture; and public behaviour becomes less predictable. This creates a plausible pathway from an information-security problem to a public-order problem.

Engineered Pathogens and Biological Weapons

Scientists, governments and AI developers are increasingly concerned that advanced AI could assist in the development of engineered pathogens and biological weapons. The concern is not limited to recreating known diseases. Experts worry that sufficiently capable systems could lower barriers to designing biological agents with combinations of traits that natural evolution rarely produces together—high transmissibility, prolonged infectiousness before symptoms appear, high rates of severe illness or death, the ability to evade existing vaccines or treatments, greater environmental stability, or efficient spread through multiple routes. Such combinations could enable widespread silent transmission before detection and produce outbreaks that are faster, harder to contain, and more lethal than the natural pandemics of the past century.

Advanced AI systems are becoming better at scientific reasoning, literature synthesis, laboratory troubleshooting, experimental planning and the use of specialised biological tools. The UK AI Security Institute reports that frontier models now match or exceed PhD-level expert performance on some relevant evaluations. The 2026 International AI Safety Report notes that general-purpose models can provide information relevant to biological and chemical weapons development, and that several frontier developers strengthened safeguards after they could no longer rule out meaningful assistance to technically capable users.

Important limitation: useful information is not the same as successfully producing and deploying a dangerous pathogen or biological weapon. Laboratories, equipment, specialist knowledge and physical materials remain necessary. The state-actor scenario is more concerning than the novice scenario. Trained scientists, laboratories and procurement capability already exist; AI may accelerate research, search, troubleshooting and experimental iteration. Frontier laboratories including Anthropic and OpenAI now explicitly treat advanced assistance with biological or chemical weapons as a catastrophic-risk category.

Potential household consequences resemble those of an unusually severe pandemic: overloaded hospitals, medical shortages, workforce absenteeism, interrupted essential services, supply-chain disruption, and periods when avoiding outside contact is highly desirable. Advanced measures for heightened biological risks include a clean-air refuge with a suitably-sized CBRN (Chemical, Biological, Radiological and Nuclear) air filtration system with positive overpressure management.

Military and Geopolitical Escalation

AI is increasingly embedded in military intelligence, autonomous systems, targeting, cyber operations, logistics and command support. Australia’s 2026 defence strategy explicitly identifies AI and autonomous systems as increasingly important to warfare. Potential future effects include faster intelligence processing, autonomous drone and vehicle systems, accelerated targeting cycles, closer coordination of cyber and kinetic operations, greater electronic-warfare capability, and machine-speed decision support.

A major-power conflict does not need to involve direct attack on Australian cities to affect households. More likely household effects may include cyberattack, disrupted trade, fuel shortages, communications problems, satellite disruption, economic shock and supply-chain interruption. In more severe conflict, blast, fallout, CBRN hazards and direct attack on strategic facilities become relevant.

From Technical Disruption to Physical Insecurity

Infrastructure failure can cascade into human-security problems. Electricity or telecommunications disruption can render payment systems unreliable, impair fuel distribution, degrade refrigeration and logistics, force businesses to close or ration goods, and place emergency services under pressure. Short disruptions produce inconvenience, queues and temporary shortages. Longer ones make scarcity more important, confidence fall, opportunistic crime more likely, response times worse, and public-order problems more plausible.

Information disruption can amplify the same process: people cannot verify warnings or shortage reports; false reports generate panic behaviour; and genuine authorities struggle to establish trust. Civil disorder is not inevitable. Australia and New Zealand have strong institutions, emergency-management capacity and generally high social stability. Yet serious continuity planning should not assume that normal police response, normal commercial supply or normal public behaviour will remain unchanged through every prolonged disruption.

Physical security therefore has a legitimate place in AI-related preparedness. As a baseline, this starts with sound doors and windows, lighting, detection, a secure internal retreat and sensible property layout. For those willing to implement more advanced physical security measures, it may extend to comprehensive hardening of the primary residence or a purpose-built hardened living structure; with greater forced-entry resistance and the ability to support normal family life during prolonged degraded security. Such measures only make sense as part of a wider continuity system covering water, energy, food, sanitation, communications and mobility.

Preparing for the Effects Rather Than the Trigger

Different AI scenarios repeatedly produce similar household requirements. Independent water helps under cyberattack, grid failure, natural disaster or prolonged service disruption. Independent power helps under grid failure, communications outages, supply disruption or broader emergencies. Food and essential supplies reduce reliance on supermarkets, functioning freight and electronic payments. Communications redundancy reduces dependence on any single network. Medical supplies and isolation capability matter more during biological events. Physical security matters more if policing is delayed, opportunistic offending increases or public order deteriorates. Serious continuity planning should therefore be capability-based rather than scenario-specific.

Keep Critical Systems Locally Controllable

Modern systems increasingly connect to vendor cloud platforms—batteries, solar inverters, vehicles, gates, cameras, pumps, access control and building automation. Connectivity provides monitoring, updates, diagnostics and convenience; it may also give external parties some degree of operational authority.

For every critical system it is worth asking: Will it continue essential operation indefinitely without internet? Does it require recurring cloud authentication? Who can remotely issue commands? Does it contain cellular connectivity? Can its WAN connection be physically removed? Can it still be locally controlled after isolation? What happens if the manufacturer goes bankrupt or its servers disappear? Can it be restored locally after malicious or accidental configuration changes?

Concrete example: if your battery system requires cloud authentication to discharge or change operating modes, test what happens when the connection is cut. The same applies to smart gates, pumps and access-control systems. ASD critical-infrastructure guidance increasingly recommends identifying vital systems, mapping their external connections, creating isolation points and testing isolated operation. The same principle can be applied at household scale, especially for power, water, ventilation, access control and communications.

Vendor jurisdiction also matters when the manufacturer retains meaningful remote authority. Australian government procurement guidance already considers foreign ownership, foreign control and foreign government influence. The sensible approach is not simply “avoid products from country X,” but to assess how much remote authority the vendor actually has. A foreign-made device with no external command path may create little geopolitical risk; a highly connected system with remote operational control deserves more scrutiny regardless of origin. Avoid unnecessary concentration: one vendor controlling battery, vehicle, cameras, access and home automation creates common-mode failure. Vendor diversity can itself be a form of redundancy.

Reduce Your Personal Digital Footprint

AI lowers the effort required to build a detailed profile of someone. Useful source material includes old social posts, photos, business biographies, property information, family accounts, leaked credentials, voice recordings and travel history. Reducing unnecessary exposure makes impersonation, social engineering, personalised manipulation and physical reconnaissance harder.

Consider limiting public disclosure of current location, travel plans, family relationships, children’s details, home address, property layout, vehicles, security arrangements and unusual continuity infrastructure. Extensive voice and video material can make synthetic impersonation easier. Complete invisibility is unrealistic; the goal is to reduce volume, freshness and connectivity of personal information.

Important accounts—email, banking, telecommunications, cloud storage and social accounts—deserve especially strong protection: unique passwords, passkeys, phishing-resistant multi-factor authentication and physical FIDO2 keys where appropriate. Email is particularly critical because it often controls account recovery for everything else. Consider separating public or business contact details from private family communications and from financial or account-recovery addresses.

Authentication in an AI World

Voice and appearance are increasingly weak forms of authentication. Families and close business partners should have simple, pre-agreed procedures for high-consequence requests: urgent money transfers, changes of payment details, unusual requests for property access, emergency pickups, travel-plan changes, or requests to disclose sensitive information.

Do not rely solely on a familiar voice, a familiar face on video, a familiar phone number, or a familiar-looking email. Verify through a known callback number, a second communications channel, or a pre-agreed family or business authentication method.

One particularly robust approach is a small set of private “proof-of-life” questions that can only be answered correctly by the real person. These should draw on shared memories that have never been recorded in emails, messages, social media, photos, or any other digital form—details that exist solely between the two people involved. Examples might include:

• What did you want to be when you grew up at age four?

• What private nickname did you give me when we first started dating?

• What was the name of the café where we had our first serious argument?

• What was the specific detail you noticed about the house the day we moved in?

  
  

Because these details live only in shared human memory, they are extremely difficult for an AI system—even one that has thoroughly analysed a person’s digital footprint—to invent or retrieve. Keep the questions few, simple, and periodically refreshed. The same principle applies to business partners for consequential financial or operational decisions: require a pre-agreed verification step that rests on non-digital, shared knowledge rather than recognition alone.

Retain the Ability to Function Without AI

Useful technologies can gradually replace underlying human competence. GPS is a clear example: navigation becomes easier while people learn fewer routes, and spatial skills deteriorate through disuse. AI may extend this effect into equipment troubleshooting, mechanical diagnosis, research, route planning, cooking, first-pass medical reasoning, financial administration, writing and interpreting documents, and crisis decision-making. AI can improve normal capability while simultaneously reducing fallback competence.

The risk appears if the internet disappears, the provider fails, the service is unavailable, the system becomes untrusted or access is restricted. The sensible response is not to reject AI, but to retain enough underlying competence to function when it is absent. Useful offline resources include paper or offline maps, first-aid references, equipment manuals, wiring diagrams, operating procedures, agricultural references, food-preservation information, important contacts and critical system configurations. More important than documents is actual practice and knowing how systems work.

Useful skills to retain include navigation, first aid, basic electrical and mechanical troubleshooting, equipment repair, food preparation, property-system operation and practical problem-solving. Ensure that more than one household member understands the power system, water system, communications, isolation procedures and key security systems.

Test the Fallback Before You Need It

Hidden dependencies often remain invisible until disconnected. Periodically test the property with grid power unavailable and external internet unavailable, and observe what stops working. Possible discoveries include batteries that require cloud access, gates that lose functionality, cameras that stop recording, pumps that depend on another controller, smart devices that cannot be locally operated, or household members who no longer remember manual procedures.

Test independent power, water, communications, access control, lighting, refrigeration and backup internet or radio. Also test human assumptions: how long water actually lasts, which foods depend on refrigeration, whether family members know emergency contacts, and whether people can navigate without phone apps.

Recovery matters as well as isolation. Disconnecting a compromised system does not undo malicious configuration, corrupted firmware or stolen credentials. Consider known-good configurations, local administrator credentials, offline backups, printed commissioning information, and spare controllers or critical components where justified.

Human and Community Resilience

Long-duration resilience is not purely technological. Local relationships can become more important when national systems are degraded. Useful networks may include neighbours, farmers, mechanics, electricians, medical professionals, tradespeople and local suppliers. A technologically sophisticated but socially isolated household may be less resilient than one embedded in a functioning local community. Practical skills, adaptability and relationships become more important as disruption length increases. Financial flexibility also matters: reduced fixed commitments, cash reserves and the ability to adapt work or income. AI-driven economic disruption may matter even without a conventional emergency.

Where Private Preparedness Stops

Households cannot meaningfully solve AI alignment, frontier-model governance, national cyber defence, biological-weapons control, nuclear escalation or international military stability. Genuine superintelligence or loss-of-control scenarios are even harder to plan around. Current systems do not possess all the capabilities generally associated with genuine loss of control. Relevant capabilities are improving, yet probability remains disputed and the form of any such threat is unknown. There is no credible basis for claiming an “AI-proof” bunker, an “AI-proof” property, a Faraday cage as a universal AI solution, or any single physical preparation as protection against superintelligence.

Private preparedness is strongest where the failure mechanisms can be identified, household consequences are plausible, and preparation improves outcomes across several scenarios.

Practical Priority Summary

High-value measures that remain useful under many ordinary emergencies as well as AI-related disruption include:

• Independent water and power supplies that function without external networks

• Food and essential medical reserves that reduce dependence on just-in-time logistics and electronic payments

• Physical cash and non-electronic payment reserves sufficient for several weeks of essential purchases

• Fuel and basic mobility readiness (rotated fuel stores, vehicle preparedness, alternative transport options where practical)

• Communications redundancy (multiple networks, offline options)

• Strong account security (unique credentials, passkeys, phishing-resistant MFA, protected recovery email)

• Reduced unnecessary digital exposure of location, family details and property information

• Local control and isolation capability for critical household systems

• Sanitation, hygiene and waste-management capability that functions when municipal services are interrupted

• Physical security measures and a secure internal retreat

• A clean-air refuge with CBRN air filtration and positive overpressure

• Printed or offline copies of critical documents, contacts, maps and system information

• Basic tools, common spare parts and the ability to perform simple repairs

• Practical human skills and documented procedures that more than one household member can execute

• Local mutual-aid relationships with neighbours, tradespeople and community contacts

• Regular testing of fallbacks under realistic disconnection conditions

These steps improve outcomes whether the trigger is cyberattack, natural disaster, pandemic, supply-chain failure, information disruption or broader systemic stress.

Closing Points

Nobody preparing today knows what frontier AI will look like in 2030, 2040 or 2050. Some feared developments may prove manageable; other important capabilities may emerge through routes receiving little attention today. The most useful preparations tend to remain valuable under many ordinary emergencies as well.

Many AI-related risks ultimately reach households through familiar conditions: infrastructure that stops working, information that cannot immediately be trusted, external systems that become unavailable or untrusted, public institutions operating under pressure, and physical security becoming more important as disruption persists. Private individuals cannot control how advanced AI develops. They can, however, reduce critical dependencies, preserve fallback options, improve their ability to operate independently, and make themselves less vulnerable when normal systems fail.

Selected sources

·       UK AI Security Institute cyber and biology evaluations and Frontier AI Trends reporting;

·       UK National Cyber Security Centre assessments of AI and cyber threat;

·       2026 International AI Safety Report;

·       Australian Signals Directorate / CI Fortify guidance on isolating vital systems;

·       Australia’s 2026 National Defence Strategy and related Innovation, Science and Technology Strategy;

·       Public statements and risk frameworks from Anthropic and OpenAI on biological and chemical risks.