Submission 566
Duntulm Partners
Submitter group: Companies.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
Duntulm Partners, a resilience adviser, recommends enforceable AI infrastructure safeguards: resilience plans, human accountability, community benefits, frontier-developer disclosures, inclusion and employment goals, and First Nations consent. It proposes sovereign design capability, distributed infrastructure and Pacific partnerships. It offers public-utility treatment as a direction to test, acknowledging investment trade-offs, and disclaims energy, water and planning expertise.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Recommends resilience plans before development approval for covered facilities, updated at gates such as energy connection approval and commissioning.
We therefore recommend that a resilience plan be lodged before development approval for covered facilities and updated at defined gates, such as energy connection approval and commissioning.
Recommends that each plan name a single accountable officer at the facility and, for training, at the AI developer.
Each plan should name one accountable officer at the facility, and for training activity one accountable officer at the AI developer, identified before approval and not reconstructed afterwards.
Recommends that each plan include a community resilience statement of the facility's net local effects.
We recommend that each plan include a short community resilience statement, signed by the accountable officer, setting out the facility's net effect on local energy, water, workforce, emergency capacity and land, how any negative effect is offset, and how the community has been invited to test it.
Recommends that flexibility obligations include a protected class for facilities serving health, emergency and utility services.
We recommend that flexibility obligations include a protected class for facilities or workloads serving health, emergency and utility services, defined in advance with the relevant authorities, and that curtailment follow a published order of priority that has been rehearsed rather than discovered during an event.
Argues developer disclosures need named accountable sign-off and competent official review, and that voluntary disclosure is insufficient.
Each disclosure should be signed by a named accountable officer at the developer and reviewed by officials with the capacity to question it, so that the obligation is not a formality. We do not think voluntary disclosure is enough for technology of this consequence.
Argues that sovereignty should be tested by whether essential services could continue without a single foreign provider.
Sovereignty is not only where a facility sits. It is whether Australia could keep essential services running, and make its own decisions, if access to a foreign model, platform or supplier were reduced or withdrawn, whether through commercial decision, geopolitical pressure, outage or attack.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Require a resilience plan to be lodged before development approval for covered facilities and updated at defined gates such as energy connection approval and commissioning.
We therefore recommend that a resilience plan be lodged before development approval for covered facilities and updated at defined gates, such as energy connection approval and commissioning.
Require pre-approval resilience planning to model compound cyber/extreme-weather events, lifetime climate and physical attack/conflict from the design stage.
Require a resilience plan before development approval, not after operations begin, and require it to model compound events such as a cyber incident during extreme weather, the climate projected over the facility's operating life, and physical attack and conflict as planning inputs from the first drawing.
Allow a transition period after operations begin for reporting and certificate obligations, but not for siting and design obligations.
A transition period after operations begin is reasonable for reporting and certificate obligations, but not for siting and design obligations.
For projects already in the pipeline, require the plan at the next approval gate not yet passed.
For projects already in the pipeline, the plan should be required at the next approval gate not yet passed.
Have resilience obligations in the plan also follow criticality, drawing on the SOCI Act definition of critical data storage or processing assets, while energy obligations follow capacity.
We suggest that resilience obligations in the plan also follow criticality, drawing on the existing definition of critical data storage or processing assets under the SOCI Act (page 5), while energy obligations follow capacity.
Require plans to show performance under concurrent stress.
We support the requirement to build operational resilience against water disruption, drought and climate change (page 14), and recommend that plans show performance under concurrent stress.
Require data centres to demonstrate that proposed water use will not harm local water security under drought and heat scenarios, not only average conditions.
Data centres should demonstrate that proposed water use will not harm local water security under drought and heat scenarios, not only under average conditions.
Make siting criteria risk-based, including bushfire, flood and heat exposure as well as proximity to schools and homes.
On minimum distances from schools and homes, we suggest siting criteria be risk-based and include bushfire, flood and heat exposure as well as proximity.
Consider streamlined approvals for facilities with independently assessed resilience-by-design plans.
Facilities with independently assessed resilience-by-design plans could be offered streamlined approvals, which rewards the investment the standards ask for and gives investors certainty that comes from durable principles.
Test efficiency, cooling capacity and water demand against hot days and drought conditions projected for the facility's operating life, not historical averages.
Efficiency, cooling capacity and water demand should instead be tested against the hot days and drought conditions projected for the facility's operating life, not historical averages.
Assess plans against forward-looking climate projections for heat, drought, bushfire, flood and storm across the whole operating life, including where energy, water and telecommunications connect.
We recommend that plans be assessed against forward-looking climate projections for heat, drought, bushfire, flood and storm across the whole operating life, including at the points where energy, water and telecommunications connect;
Disclose design margins for cooling, water and power, and refresh the climate assessment on a fixed cycle and whenever projections change.
We recommend that plans be assessed against forward-looking climate projections for heat, drought, bushfire, flood and storm across the whole operating life, including at the points where energy, water and telecommunications connect; that design margins for cooling, water and power be disclosed; and that the assessment be refreshed on a fixed cycle and whenever projections change.
Avoid siting in locations where exposure is rising unless the risk can be demonstrably managed.
Siting should avoid locations where exposure is rising unless the risk can be demonstrably managed.
Let developers authorised to train in Australia schedule, defer and shift training to when renewable supply is abundant and avoid adding load during heat and grid stress events.
developers authorised to train in Australia should be able to schedule, defer and shift training to when renewable supply is abundant, and to avoid adding load during heat and grid stress events.
Treat physical attack and conflict as foreseeable risks alongside natural hazards in Australian planning of data centres.
Australian planning should treat physical attack and conflict as foreseeable risks alongside natural hazards.
Name one accountable officer at the facility, and for training activity one at the AI developer, identified before approval.
Each plan should name one accountable officer at the facility, and for training activity one accountable officer at the AI developer, identified before approval and not reconstructed afterwards.
Write the plan down, own it in a role rather than a person, and review it annually and after any major change.
The plan should be written down, owned in a role rather than a person, and reviewed annually and after any major change, so that coordination does not leave when individuals do.
Give regulators who review plans the authority, competence, time and information to disagree with what they are shown.
Regulators who review plans need the authority, competence, time and information to disagree with what they are shown.
Adopt a single national reporting template, in plain language, signed by the named accountable officer, published and aligned with NGER Act reporting.
On reporting (page 24), we recommend a single national template, written in plain language, signed by the named accountable officer and published, aligned with emissions and energy reporting under the NGER Act reforms (page 6) to avoid duplication.
Reconcile demand flexibility and backup-generation limits with essential-service continuity, and require baseline security for new connected generation and storage under renewable obligations.
Reconcile demand flexibility and backup generation limits with continuity for the essential services these facilities host, and require the new connected generation and storage brought forward by the renewable obligation to meet baseline security standards.
Define protected health, emergency and utility facilities or workloads with relevant authorities in advance; require published, rehearsed curtailment priorities.
We recommend that flexibility obligations include a protected class for facilities or workloads serving health, emergency and utility services, defined in advance with the relevant authorities, and that curtailment follow a published order of priority that has been rehearsed rather than discovered during an event.
Test plans in exercises that combine a grid stress event with a cyber incident.
Plans should be tested in exercises that combine a grid stress event with a cyber incident.
Require connected capacity counted towards the electricity obligation, including on-site generation and batteries, to meet baseline security and recoverability requirements at procurement and installation.
an obligation designed around certificates can reward the volume of connected assets without regard to their security. Capacity counted towards the obligation, including on-site generation and batteries, should meet baseline security and recoverability requirements at procurement and installation.
Require signed community resilience statements covering net local energy, water, workforce, emergency-capacity and land impacts, offsets for negative effects, and how the community was invited to test them.
We recommend that each plan include a short community resilience statement, signed by the accountable officer, setting out the facility's net effect on local energy, water, workforce, emergency capacity and land, how any negative effect is offset, and how the community has been invited to test it.
Extend no-net-cost requirements to emergency plans with local brigades and emergency services before approval, pre-incident familiarisation and site access, and hazard-specific training and equipment funding.
recommend that the no net costs principle expressly cover emergency management: a facility emergency plan developed with local brigades and emergency services before approval, pre-incident familiarisation and site access, and funding for training and equipment where the facility creates a specific hazard.
Expect community engagement led by named senior people from the proponent who are present in the community, not only through portals and templated communications.
First, engagement led by named senior people from the proponent who are present in the community, not only through portals and templated communications.
Expect a published record of what the community said and what changed as a result, with unresolved concerns recorded as raised.
Second, a published record of what the community said and what changed as a result, with unresolved concerns recorded as they were raised, not smoothed.
Expect early engagement with councils, First Nations peoples, local businesses and local emergency services.
Third, early engagement with councils, First Nations peoples, local businesses and local emergency services.
Expect disclosure where AI has been used to draft or summarise engagement material, and a named person who has read the unfiltered feedback.
Fourth, disclosure where AI has been used to draft or summarise engagement material, and a named person who has read the unfiltered feedback.
Have a named person read dissenting and unusual consultation submissions in full when AI summarisation is used.
We note the Government's statement that AI tools may be used to summarise submissions to this consultation (page 25) and do not object. We ask only that the same principle apply to everyone: summarisation tends towards the average, so a named person should read the dissenting and the unusual in full.
Have operators fund pathways in cyber security and operational resilience, not only trades.
The paper lists cyber security specialists among occupations in short supply (page 19), so operators should fund pathways in cyber security and operational resilience, not only trades.
Train staff who operate or oversee AI-enabled critical systems in meaningful oversight, including recognising automation bias.
Staff who operate or oversee AI-enabled critical systems should be trained in meaningful oversight, including recognising automation bias, which peaks under fatigue and time pressure, as in an operations centre during an incident.
Focus skills development on local workforces as far as possible, rather than temporary fly-in workers.
Skills and capability development should focus on local workforces as far as possible, not on a temporary fly-in workforce.
Require measurable local targets, such as host-region apprenticeship and traineeship hours, TAFE and training-provider partnerships, local procurement and underrepresented-cohort pathways, reporting progress in community resilience statements.
We recommend that operators and developers set measurable local targets, such as apprenticeship and traineeship hours drawn from the host region, partnerships with local TAFEs and training providers, local procurement, and pathways for underrepresented cohorts in the host community, and that progress be reported in the community resilience statement.
Include in training skills a community needs in a disaster, such as electrical, HVAC, communications, climate adaptation and emergency management.
Training should include skills a community needs in a disaster, such as electrical, HVAC, communications, climate adaptation and emergency management capability, so that the workforce built for the facility also strengthens local recovery capacity.
Pair not-for-profit access to compute with data-protection, cyber-security and AI-oversight support, potentially funded through public-interest research payments.
Local capability, skills and research. The paper rightly lists not-for-profit organisations among those who should benefit from access to compute (page 22). Our work with mission-driven organisations shows that many hold sensitive information about vulnerable people and lack the governance and cyber capacity to use powerful tools safely. Access to compute should therefore be paired with support for data protection, cyber security and AI oversight, and public-interest research payments could fund it.
Measure human outcomes as well as inputs: working oversight in not-for-profits and agencies, and graduates in Australian jobs and, where possible, host regions.
Indicators should measure outcomes in human terms, such as the number of not-for-profits and public agencies with working oversight arrangements and graduates employed in Australian roles and, where possible, in the regions hosting the facilities, as well as inputs such as dollars and partnerships.
Have developers support AI literacy for decision-makers such as boards, executives and public servants.
we suggest developers support AI literacy for decision-makers such as boards, executives and public servants.
Ground AI regulatory frameworks and standards in a robust ethical framework that is Australian in origin and content, not imported.
Any regulatory framework or standards for the development, use and implementation of AI should be grounded in a robust ethical framework that is Australian in origin and content, not imported.
Develop and publish Australian ethics with First Nations, multicultural, disabled, regional, worker and civil-society input, considering the Pacific context for Australia and Oceania.
We recommend that the Government develop and publish this framework with broad community input, including First Nations peoples, multicultural communities, people with disability, regional communities, workers and civil society, and with attention to the Pacific context so that it is appropriate for Australia and Oceania.
Embed accountability at every stage of the AI lifecycle, with a named accountable person at each stage and not only at the point of use.
Accountability should be embedded at every stage of the AI lifecycle, from design and data collection through training, testing, deployment, monitoring and retirement, with a named accountable person at each stage and not only at the point of use.
Apply common ethical principles across Commonwealth, state and territory law, national minimum standards and higher jurisdictional requirements.
This grounding should then underpin the legislative and regulatory framework applied by the Commonwealth and by states and territories, so that the same principles carry through the national minimum standards and any higher requirements set by individual jurisdictions (page 6), instead of each layer inventing its own.
Make infrastructure and training standards consistent with, and not a substitute for, a legislated standard of meaningful oversight and named accountability for consequential government decisions.
the standards for infrastructure and training should be consistent with, and not a substitute for, a legislated standard of meaningful oversight and named accountability for consequential government decisions.
Align data centre resilience obligations with the SOCI Act's positive security obligations to avoid duplicate reporting.
Resilience obligations for data centres should also be aligned with the SOCI Act's positive security obligations to avoid duplicate reporting.
Require relevant First Nations communities’ free, prior and informed consent before cultural-knowledge use in training, fine-tuning or model development, consistent with UNDRIP and Indigenous data sovereignty.
AI training and development must respect the rights of First Nations peoples to their cultural knowledge. Language, stories, art, ecological and cultural knowledge, and data about communities belong to the peoples who hold them. No such knowledge should be used or incorporated in AI training, fine-tuning or model development without the free, prior and informed consent of the relevant First Nations communities, consistent with the United Nations Declaration on the Rights of Indigenous Peoples, which Australia has endorsed, and with the principles of Indigenous data sovereignty.
Require prior consent for First Nations knowledge use, explaining the use and rights to refuse or withdraw; recognise collective ownership and sacred or restricted knowledge.
No such knowledge should be used or incorporated in AI training, fine-tuning or model development without the free, prior and informed consent of the relevant First Nations communities, consistent with the United Nations Declaration on the Rights of Indigenous Peoples, which Australia has endorsed, and with the principles of Indigenous data sovereignty. Consent must be given before use, with a clear explanation of how the knowledge will be used, the right to refuse or to withdraw, and recognition that knowledge may be held collectively and that some of it is sacred or restricted.
Require demonstrated consent practices, exclusion of unconsented First Nations content, fair benefits for participating communities and First Nations-led data/model governance.
We recommend that conditions for AI training require developers to demonstrate how they obtain and honour that consent, to identify and exclude First Nations content collected without it, to share benefits fairly with communities who choose to participate, and to support First Nations-led governance of any data and models built with that knowledge.
Co-design First Nations training-consent safeguards with First Nations peoples and connect them to the Attorney-General’s copyright work and existing Indigenous cultural and intellectual property protections.
This should be designed with First Nations peoples and not for them and should connect with the Attorney- General's copyright work and with existing protections for Indigenous cultural and intellectual property.
Anchor Australian frontier-training conditions in durable principles, meaningful human oversight and named accountability rather than specifications overtaken by model changes.
We do not suggest importing the EU architecture. We suggest anchoring Australian conditions in durable principles, meaningful human oversight and named accountability, rather than in technical specifications that the next generation of models will overtake.
Include in reportable incidents harms in deployment where over-reliance on model output contributed, not only security breaches and capability failures.
Reportable incidents should include harms in deployment where over-reliance on model output contributed, not only security breaches and capability failures.
Have developers authorised to train in Australia give the AI Safety Institute their evaluation results on sycophancy and over-reliance.
Developers authorised to train in Australia should provide the Australian AI Safety Institute with the results of their evaluations of sycophancy, meaning the tendency to favour agreeable and validating answers over accurate ones, and of over-reliance.
Require developers authorised to train in Australia to give deployers plain-language guidance on foreseeable over-reliance and oversight that works under time pressure.
Developers authorised to train in Australia should provide the Australian AI Safety Institute with the results of their evaluations of sycophancy, meaning the tendency to favour agreeable and validating answers over accurate ones, and of over-reliance. They should also give deployers plain-language guidance on foreseeable over-reliance and on designing oversight that works under time pressure.
Have the legal obligation to inform government govern what is provided proactively, held ready on request and published, with safeguards protecting security-sensitive and commercially sensitive detail.
The legal obligation set out below should govern what is provided proactively, what is held ready on request and what is published. Safeguards should protect security-sensitive and commercially sensitive detail.
Add human-AI interaction to Institute research: high-stakes automation bias and habitual reliance’s long-term effects on skill and judgement, treating those long-term effects as an open question.
Research. The Institute's early priorities include model alignment and multi-agent risk (page 21). We suggest adding human and AI interaction, including automation bias in high-stakes settings and the long-term effect on skill and judgement of habitual reliance on AI. This last point is an open question, not a settled conclusion:
Require frontier developers authorised to train in Australia to disclose high-level training data, objectives and evaluation methodology to government at authorisation, major training runs and releases, without exposing trade secrets.
we recommend that frontier developers authorised to train in Australia be legally obliged to inform the Australian Government, proactively and at defined points such as authorisation, major training runs and major releases, about: their high-level training methodology, including the kinds and sources of data used, the objectives set and the evaluation approach, at a level that serves oversight without exposing trade secrets; how they comply with a digital duty of care, building on the Government's announced priority (page 21); their non-discriminatory practices, including the bias risk assessments and testing described below; and
Require frontier developers authorised to train in Australia to disclose digital-duty-of-care compliance to government at authorisation, major training runs and major releases.
we recommend that frontier developers authorised to train in Australia be legally obliged to inform the Australian Government, proactively and at defined points such as authorisation, major training runs and major releases, about: their high-level training methodology, including the kinds and sources of data used, the objectives set and the evaluation approach, at a level that serves oversight without exposing trade secrets; how they comply with a digital duty of care, building on the Government's announced priority (page 21); their non-discriminatory practices, including the bias risk assessments and testing described below; and
Require frontier developers authorised to train in Australia to disclose non-discriminatory practices, bias risk assessments and testing to government at authorisation, major training runs and major releases.
we recommend that frontier developers authorised to train in Australia be legally obliged to inform the Australian Government, proactively and at defined points such as authorisation, major training runs and major releases, about: their high-level training methodology, including the kinds and sources of data used, the objectives set and the evaluation approach, at a level that serves oversight without exposing trade secrets; how they comply with a digital duty of care, building on the Government's announced priority (page 21); their non-discriminatory practices, including the bias risk assessments and testing described below; and
Require frontier developers authorised to train in Australia to disclose safety evaluations, controls, incidents and training energy/water use to government at authorisation, major training runs and major releases.
frontier developers authorised to train in Australia be legally obliged to inform the Australian Government, proactively and at defined points such as authorisation, major training runs and major releases, about: their high-level training methodology, including the kinds and sources of data used, the objectives set and the evaluation approach, at a level that serves oversight without exposing trade secrets; how they comply with a digital duty of care, building on the Government's announced priority (page 21); their non-discriminatory practices, including the bias risk assessments and testing described below; and safety and security considerations, including evaluation results, security controls and reportable incidents, together with the energy and water used by training
Include assessment against projected climate impacts in training safety and security considerations.
safety and security considerations, including evaluation results, security controls and reportable incidents, together with the energy and water used by training and how the activity has been assessed against projected climate impacts.
Have each disclosure signed by a named accountable officer at the developer and reviewed by officials with the capacity to question it.
Each disclosure should be signed by a named accountable officer at the developer and reviewed by officials with the capacity to question it, so that the obligation is not a formality.
Reward, then require, sovereign training design: Australian design/evaluation staff, data, languages and contexts; community-controlled First Nations knowledge; local evaluation, adaptation and fine-tuning against Australian values/law; transparency on decision locations and decision-makers.
We recommend that conditions for large-scale training reward, and over time require, genuine sovereign design capability: Australian researchers and engineers involved in design and evaluation decisions and not only in operations; Australian data, languages and contexts represented in training and evaluation, including First Nations knowledge held under the control of the relevant communities; the ability to evaluate, adapt and fine-tune models locally against Australian values and law; and transparency about where design decisions are made, and by whom.
Require training conditions to include an inclusive approach designed to minimise or eliminate bias on discriminatory grounds such as age, race, religion, sex and disability.
Conditions for training should require an inclusive approach explicitly designed to minimise or eliminate bias on discriminatory grounds such as age, race, religion, sex and disability.
Require pre-training and major-release bias assessments, representative data with disclosed gaps, cross-group testing including independent Australian evaluators, affected-community definitions of fairness, and Institute reporting alongside sycophancy and over-reliance evaluations.
We suggest five elements: a documented bias risk assessment before training begins and at each major release; representative training data, with gaps identified and disclosed; testing of outputs across groups, including by independent Australian evaluators; involvement of affected communities in defining what fair means in Australian settings; and reporting of results to the AI Safety Institute alongside the sycophancy and over- reliance evaluations proposed above.
Make findings of material discriminatory bias reportable incidents.
Findings of material discriminatory bias should be reportable incidents.
Align bias requirements with Commonwealth, state and territory anti-discrimination law.
These requirements should align with Commonwealth, state and territory anti-discrimination law, which already prohibits discrimination on many of these grounds in areas such as employment, education and services, so that automated systems are held to the standard people already are.
Require each developer to maintain a published social impact model with measurable goals for positive social outcomes and digital inclusion.
In addition, each developer should maintain a social impact model: a published plan with measurable goals showing how its models are designed to deliver positive social outcomes, not only technological improvement, and to advance digital inclusion, meaning that everyone can benefit from AI regardless of where they live, their income, age, ability or background.
Measure inclusion through access, affordability, digital ability and AI scrutiny; reach regional and remote people, disabled and older Australians, First Nations communities, low-income people, not-for-profits and small organisations.
We suggest developers measure it on the dimensions already used to track digital inclusion in Australia, which are access, affordability and digital ability, extended to include the ability to question and oversee AI. Those who must be reached include regional and remote Australians, people with disability, older Australians, First Nations communities, people on low incomes, not-for-profits and small organisations.
Require specific, measurable inclusion goals, for example on access, affordability, accessibility, language and group performance and social outcomes; report annually with independent verification, penalties and regulatory enforcement.
Goals should be specific and measurable, for example access and affordability for underserved groups, accessibility standards, performance across languages and groups, and outcomes in areas such as health, education, public services and climate resilience. They should be reported annually, with independent verification, and backed by penalties and a regulator able to enforce them.
Protect commercially and security-sensitive detail in social impact reporting, with confidentiality for what is provided to government and a plain-language public summary.
Safeguards should protect commercially and security-sensitive detail, with confidentiality for what is provided to government and a plain- language public summary.
Have the social impact model show how AI development will be a net contributor to employment and the economy in Australia, not only a means of eliminating valuable work.
The social impact model should also show how AI development will be a net contributor to employment and the economy in Australia, and not only a means of eliminating valuable work.
Have developers set out and report on the expected effects of their models on employment in Australia, by sector and region where possible.
We recommend that developers set out, and report against, four things: the expected effects of their models on employment in Australia, by sector and region where possible; the jobs, skills pathways and local businesses their activity supports; the measures they take with employers, unions and training providers to support workers whose roles change, including retraining and redeployment; and how the economic value created is shared with Australia, for example through local employment, research and procurement.
Require reporting on supported jobs, skills pathways and local businesses; retraining and redeployment with employers, unions and providers; and Australian value sharing through employment, research and procurement.
We recommend that developers set out, and report against, four things: the expected effects of their models on employment in Australia, by sector and region where possible; the jobs, skills pathways and local businesses their activity supports; the measures they take with employers, unions and training providers to support workers whose roles change, including retraining and redeployment; and how the economic value created is shared with Australia, for example through local employment, research and procurement.
Where a model is designed to substitute for people rather than support them, have the plan say so and explain how the transition will be managed.
Where a model is designed to substitute for people rather than support them, the plan should say so and explain how the transition will be managed.
Weigh retained local skills, reserved public-interest models/compute, open, portable or locally operable essential-service options, and AI service continuity in standards and frontier-training conditions.
We recommend that the standards and the conditions for frontier training weigh four things: local engineering and operating skills that remain when a foreign partner leaves; models and compute reserved for Australian public-interest use (page 23); open, portable or locally operable options for essential public services, so that no critical service depends on a single foreign provider; and continuity arrangements for critical services that depend on AI. Sovereign capability should be built with trusted partners and not in isolation, but dependence should be a choice Australia makes knowingly and can reverse.
Build sovereign capability with trusted partners while making foreign dependence a conscious, reversible choice.
Sovereign capability should be built with trusted partners and not in isolation, but dependence should be a choice Australia makes knowingly and can reverse.
Assess AI standards against cyber security, safeguarding and resilience; appoint a coordinating body reporting combined effects; integrate infrastructure, hardware, software and governance planning with local government, utilities and civil society.
We recommend that the AI standards be assessed explicitly against cyber security, safeguarding and cyber resilience; that a named body hold a standing mandate to coordinate across those areas and report on their combined effect, as our climate briefing proposed for compound resilience; and that physical infrastructure, hardware, software and governance be planned as one system and not four projects. Local government, utilities and civil society, who are closest to the people affected, should be at that table.
Position Australian AI development and infrastructure as a regional good for Oceania, built with Pacific partners and not simply for them.
Australian AI development and infrastructure should be positioned as a regional good for Oceania, built with Pacific partners and not simply for them.
Treat as design considerations from the start regional redundancy so critical Pacific services can continue if local infrastructure is lost, governance co-designed with Pacific governments and communities respecting their data sovereignty, inclusion measures extending to the region, and building Pacific capability to avoid new dependence.
We recommend that Australia treat these as design considerations from the start: regional redundancy as part of a distributed approach, so that critical Pacific services can continue if local infrastructure is lost to disaster; governance co-designed with Pacific governments and communities, including respect for their data sovereignty; inclusion measures in the social impact model that extend to the region; and a deliberate effort to avoid creating a new form of dependence, by building Pacific capability and not only supplying services.
Consider utility treatment for essential-service compute and models: equitable universal access, reliability, continuity, fair transparent pricing, public accountability and no costs shifted onto communities.
We invite the Government to consider whether foundational AI capacity, the compute and models on which essential services increasingly rely, should over time be treated like other utilities. Utility thinking brings obligations that market provision alone does not guarantee: universal and equitable access, reliability and continuity, fair and transparent pricing, public accountability, and costs that are not shifted onto communities.
Test a public-interest compute reserve with service standards for essential services and community organisations, recognising investment-attraction trade-offs rather than prescribing a settled design.
We suggest extending that logic to a public-interest compute reserve that essential services and community organisations can rely on, with service standards, rather than leaving access to depend on negotiated goodwill. We recognise the trade-offs with investment attraction and offer this as a direction to test, not a settled design.
Encourage smaller distributed inference/regional facilities near communities and renewables, alongside large training campuses; consider community ownership or benefit sharing, local workforce and shared energy/connectivity upgrades.
We suggest the Government encourage, alongside large training campuses, a distributed model of smaller facilities sited near communities and renewable generation, serving inference and regional needs, with options for community ownership or benefit-sharing, local workforce, and shared use of the energy and connectivity upgrades they require. Frontier training may still need large sites, and we do not suggest otherwise.
Require critical-service continuity after single-site loss through geographic dispersion, tested failover and appropriate trusted-allied backups, balanced against sovereignty.
We recommend that plans for facilities hosting critical services show how those services would continue if any single site were lost, through geographic dispersion, tested failover and, where appropriate, trusted allied locations for backup, balanced against the sovereignty considerations above.
Handle detailed vulnerability assessments through existing security channels rather than publishing them.
Detailed vulnerability assessments should be handled through existing security channels and not published.
Apply an aggregation rule so facilities under common ownership or control count together, with baseline resilience and security expectations applying proportionately to smaller facilities.
We recommend an aggregation rule so that facilities under common ownership or control count together, as the paper already proposes for cumulative portfolios of 150MW, and that baseline resilience and security expectations apply to smaller facilities in proportion.
Require large-facility and critical-service resilience plans to show emergency isolation: tenant/operational-technology segmentation, minimal authenticated cuttable remote management, and containment of multi-site vendor or platform access.
We recommend that resilience plans for large facilities, and for any facility hosting critical services, show how the site, a hall or a tenant can be isolated in an emergency. That means segmentation between operational technology (power, cooling, building controls) and information systems, and between tenants; remote management paths that are minimal, strongly authenticated and able to be cut; no single vendor or management platform able to reach many sites at once without a way to stop it; the ability to run in a degraded, manual or stand-alone mode; and exercises with emergency services and regulators to prove all of this works.
For large facilities and any hosting critical services, require degraded, manual or stand-alone operation and exercises with emergency services and regulators proving isolation works.
We recommend that resilience plans for large facilities, and for any facility hosting critical services, show how the site, a hall or a tenant can be isolated in an emergency. That means segmentation between operational technology (power, cooling, building controls) and information systems, and between tenants; remote management paths that are minimal, strongly authenticated and able to be cut; no single vendor or management platform able to reach many sites at once without a way to stop it; the ability to run in a degraded, manual or stand-alone mode; and exercises with emergency services and regulators to prove all of this works.
Enable operators to suspend, roll back or isolate a model or automated agent that is behaving unsafely.
operators should be able to suspend, roll back or isolate a model or automated agent that is behaving unsafely.
Use existing Security of Critical Infrastructure Act channels for reporting.
Reporting should use the channels that already exist under the Security of Critical Infrastructure Act.