Welcome to the latest edition of ASPIās Cyber & Tech Digest.
Each week, ASPI curates and contextualises the most important developments in cyber, technology, and geopolitics ā highlighting what matters and why.
This edition covers the period: 14 August to 21 August 2026.
Follow the Australian Strategic Policy Institute on Bluesky, LinkedIn, and X.
What Weāre Tracking
Our analysts unpack the weekās most consequential stories ā what happened, why it matters, and their view.
Australiaās compute leverage meets the open-weights question
What happened: Assistant Minister Andrew Charlton used the ANU Crawford School Public Policy Annual Lecture on Tuesday to argue that Australia should use its data centre boom to climb the AI value chain, and revealed that he and Industry Minister Tim Ayres are working to convert a voluntary expectation on hyperscalers into a requirement. The expectation in question is the fifth in the governmentās Expectations of data centres and AI infrastructure developers, released in March, which asks large-scale compute providers to give Australian start-ups, small businesses, researchers and not-for-profits access to compute on favourable terms, deploy engineers and researchers locally, and invest in Australian supply chains. The expectations are not legally binding. The speech followed a two-hour appearance on the Joe Walker Podcast on Sunday.
Why weāre tracking this: Taken together, the speech and interview are starting to reveal a clearer policy framework. A data centre build-out creates leverage over hyperscalers; that leverage is converted into guaranteed local compute and research capability; and that capability is used to support Australian model development rather than simply renting intelligence from abroad. What makes that realistic is a shift in what Charlton means by an Australian model. He is not talking about building an Australian ChatGPT from scratch. He is talking about taking a model someone else has already built and published openly, adapting it, and running it on Australian soil, which is the route Korea and Singapore have taken.
What people are saying:
āAround 5Ā¢ buys electricity. Around 10Ā¢ buys the physical infrastructure of the data centreā¦ Around 35ā45Ā¢ buys the computing equipmentā¦ The remaining 40ā50Ā¢ supports everything further up the stackā ā Andrew Charlton, ANU Crawford School Public Policy Annual Lecture
Australia should be āin the mix to genuinely capture value across 80 per cent of the AI stack rather than being confined to 20 per centā ā Andrew Charlton, ANU Crawford School Public Policy Annual Lecture
āThe strongest open models now sit within distance of the frontier at roughly a third of the price. Techniques that were research curiosities eighteen months ago ā¦ mean that a mid-sized organisation could now take a strong open model and make it better than a frontier system at a narrow, valuable task, using its own proprietary dataā ā Andrew Charlton, ANU Crawford School Public Policy Annual Lecture
My view: Charltonās AI economic roadmap is the most detailed yet from the government and very welcome. Heās right that intellectual property such as models and the way they underpin applications lie at the heart of the AI stack and therefore of the value chain. This is where Australia wants to play, not just at the lower end in infrastructure.
Thereās a significant policy conclusion to draw from this, as well as a note of caution. The policy is that Australia should rapidly build skills and expertise in models. For all the uncertainty about how exactly AI capabilities will develop over coming years, a very sure bet is that skills at the high end of the value chain will be useful, whether to finetune existing models or build from foundations up. The caution needed is on open weight models as a key part of the roadmap. The best OW models are Chinese, posing security risks. The commercial and geopolitical logic of OW will also continue to be tested, meaning bets on these models must be made carefully.
ā David Wroe, AI & Security Program
āāāā
Already read Tuesdayās Early Edition? What Weāre Tracking and everything under New since Tuesday is fresh. Tuesdayās stories are included again at the end so the free Friday Digest remains complete.
New since Tuesday
What Weāre Watching: a weekly scan of notable developments weāre tracking across technology, policy and geopolitics.
š Global policy
š¦šŗ Australia
Australia has passed the News Bargaining Incentive, which imposes a levy equal to 2.5 per cent of Australian advertising revenue on large search and social-media platforms unless they make qualifying deals with local publishers. Reuters reports the scheme covers Meta, Google, TikTok and LinkedIn where local advertising revenue exceeds A$250 million; platforms can avoid the levy by reaching deals with at least eight publishers, with larger offsets for spending with small and medium-sized outlets.
The Reserve Bank has named the global AI boom as a driver of Australian inflation. Deputy governor: āthereās an extraordinary boom going on in the global economy in AI and the tech sector, and particularly in trading partners that we have. Anyone whoās tried to buy a laptop recently will know that the price of memory for example has gone up sharply. That threatens inflation from overseasā (Startup Daily). Separately, Taiwan is paying every citizen NT$10,000 (about A$440), which the government attributes to the economic gains from demand for AI hardware (Focus Taiwan).
The NSW government has published data-centre planning guidelines committing the state to keep its part of development-application assessments within 75 days. Proponents are expected to show they can reduce grid-supplied electricity demand by 25 per cent of forecast average load for up to two hours, prioritise recycled water for cooling and support additional renewable generation, storage and rainfall-independent water.
Authorities have linked the Origin Energy breach, which exposed data on about 900,000 current and former customers, to a former Accenture employee in Manila. Accenture operates call centres for Origin, and ABC News reports the former employee sought to extort the company. The Australian Federal Police investigation continues.
Liverpool City Council has deployed cameras and AI at the Whitlam Leisure Centre to alert lifeguards to swimmers in difficulty. Solicitor Avinash Singh told news.com.au the system may breach the NSW Surveillance Devices Act because it tracks people without explicit consent; the council says signage provides consent.
Australian competition authorities have raided WiseTech Globalās office seeking documents and data as part of an investigation into alleged competition-law breaches. Bloomberg reports the company disclosed the raid and said it was cooperating.
šŗšø United States
With sweeping crypto legislation stalled in Congress, the US Securities and Exchange Commission and Commodity Futures Trading Commission are moving to establish rules through agency action. The SEC is set to exempt some token offerings from securities laws, Reuters reports.
Google quietly launched a program allowing verified candidates, political parties and political action committees to bypass Gmailās spam filters more easily. The New York Times reports the change follows years of Republican complaints that the filters were biased; a lawsuit making that claim was dismissed.
The US Senate Republican campaign arm has warned AI companies that opposition to data centres could cost the party a pivotal Senate seat in Ohio. A private memo obtained by Axios asks the companies to explain who benefits from data-centre development, who pays and why communities should accept the projects.
šŖšŗ European Union
The European Commission is increasingly enforcing the Digital Markets Act through technical compliance orders and regulatory dialogue rather than headline fines. Politico Europe reports formal enforcement decisions have fallen even as decisions and rulings have risen, while cases against Google are closing and Apple expects its largest outstanding case to end without another penalty. The Commission says the DMA is meant to change conduct rather than maximise fines; smaller European firms argue the quieter approach risks leaving platform market power intact and is pushing more disputes into national courts.
āāāā
š Strategic competition
China is restricting or delaying shipments to Taiwan of germanium- and quartz-based materials used in optics, semiconductors and defence systems, as well as permanent magnets used in aerospace and robotics. Taiwanese manufacturers told Nikkei Asia that prolonged customs checks have extended production lead times and cost orders; Taiwan is supporting a domestic rare-earth test line in response.
Government-backed robot-training centres have become a major source of demand for Chinaās humanoid manufacturers, buying robots, producing teleoperation data and selling that data back to the manufacturers. The Financial Times reports more than 90 centres were operating or being built by June, but investors and analysts warn the circular, policy-supported model obscures weak commercial demand and produces data that may not transfer well to real factories and warehouses.
Unitreeās mass-market robot dogs closely follow locomotion and actuator designs developed through openly published US Army-funded research, according to researchers involved in the original work. Reuters reports the researchers do not advocate secrecy; they say the strategic failure was the United Statesā inability to commercialise and manufacture the technology at scale while Chinaās industrial system did so. Unitree was added to the Pentagonās list of Chinese military companies in June, and its robots have appeared in Chinese military exercises.
Hangzhou has deployed 15 wheeled humanoid robots to identify traffic violations, issue warnings, direct road users and answer questions. Developer SUPCON told Reuters that roughly 50 robots now operate across eight Chinese cities and that deployments could reach 200 by year-end; the machines use cameras and radar, while prospective overseas sales would face integration, certification and data-protection barriers.
Chinese AI companies can legally access restricted high-end Nvidia computing power through data centres in Southeast Asia because US export controls govern ownership and physical transfer of chips, not remote cloud use. CNBC reports the proposed Remote Access Security Act would give Washington authority to regulate that access, but implementing it would still require enforceable rules defining covered compute, prohibited customers and cloud-provider identity checks.
China has allowed ByteDance and Tencent to receive about 10,000 Nvidia H200 chips each in recent weeks, Financial Times reports. US export licences permit sales of the processors to mainland China and Hong Kong.
āāāā
š§ AI models, agents & compute
Stripe is acquiring OpenRouter for a reported US$7.5 billion, The New York Times reports, with US$1.5bn going to the startupās founders and US$6bn to its investors. OpenRouter says it will continue under the same name, product and roadmap. OpenRouter routes developer requests across multiple model providers, allowing applications to switch between frontier and open-weight models without re-engineering.
Ant Groupās inclusionAI released the full Ling-3.0 open-weight family to Hugging Face on 19 August ā tiny and flash in base, -30T and -midtrain variants, and int4 and fp8 quantisations. The int4 and fp8 quantisations are built to run on lower-specification hardware than the full-precision models require.
Former US State Department AI diplomat Daniel Remler argues the United States should test, standardise and selectively restrict Chinese open-weight models rather than ban or ignore them. Writing in Just Security, he cites US government evaluations finding weak safeguards and ideological alignment, and says locally hosting Chinese models removes data-flow risks but not the modelsā embedded political biases.
OpenAI has paused some deployment-focused training and kept its largest planned frontier training run on hold while it rewrites its safety framework. Axios reports the changes followed evidence that its unreleased Astra model may have reached the companyās critical threshold for cyber capability and that another unreleased model may have been used in the recent Hugging Face breach.
The company is also testing Private Safety Processing, a system designed to detect misuse patterns across related interactions while preserving zero-data-retention protections for eligible enterprise and API customers. Axios reports the approach is being trialled with early customers.
āāāā
šµļø Surveillance states
Flock Safety has built an AI system for police that can combine licence-plate scans with arrest records, case files, dispatch logs and commercial identity data. Code examined by Wired shows its prewritten searches can identify potential witnesses from driving patterns, infer a vehicleās associates and generate dossiers without officers beginning with a named suspect or known plate. Flock says OS Investigate remains in development with a small group of law-enforcement partners.
āāāā
āļø Platform accountability
Research by the Anti-Corruption Data Collective identified 152 Polymarket wallets that made rapid, long-shot bets on military and defence events with an average 97.2 per cent win rate, raising the possibility that some traders were using inside military information. Reuters reports some of those wagers appeared to attract large copycat bets and automated trading before US military action, potentially turning suspicious market activity into a public signal for foreign intelligence services. The researchers acknowledge luck and other explanations are possible, but argue the highest-risk military markets should be prohibited.
A study of Xās recommendation system found that replies to posts carrying values users opposed could drive the platform to serve them more of the same material. 404 Media reports the observational study covered 715 US users and found the feedback loop was especially pronounced among participants who identified as Democrats; the researchers said further work was needed to explain that difference.
Meta ran advertisements for an AI tool that encouraged users to create non-consensual nude images of female politicians. One advertisement included a pornographic deepfake resembling a prominent US politician, and Apple removed the app from the App Store after an inquiry from Wired.
Apple will move developers distributing apps in the European Union to unified business terms from 1 October, following collaboration with the European Commission. The new model replaces the per-install Core Technology Fee with a 5 per cent commission on digital transactions in apps distributed outside the App Store, permits Apple in-app purchases alongside alternative payments and changes commission tiers.
āāāā
š§ Online harms & child safety
Roblox has signed a court-enforceable undertaking with the eSafety Commissioner after roughly two years of pressure from the regulator. The platform has committed to preventing adults from contacting unknown children without parental consent, making childrenās account settings private by default, building reporting mechanisms that notify users of complaint outcomes, and engaging an independent third-party auditor to assess its safety measures, including its age-estimation deployment. A court-enforceable undertaking can be enforced in the Federal Court, unlike the industry codes eSafety has otherwise relied on.
Two US senators have demanded records and answers from TikTok about an experiment that disabled an algorithmic safety measure for 10 per cent of US users, including minors. A confidential company document said the control group existed to measure the safeguardās effect on daily active users and other core metrics; one selected user was a 16-year-old whose account received thousands of videos about sadness, hopelessness and suicide before he died. Senators Marsha Blackburn and Richard Blumenthal have asked TikTok to identify every similar experiment and explain why minors were included by 1 September.
OpenAI has launched ChatGPT for Teens, a version for users aged 13 to 17 with stronger protections around suicide, self-harm and sexual or romantic conversations, as well as study tools. The service requires parental permission and gives parents controls over settings and usage.
Former Meta engineering director Arturo BĆ©jar has testified that Mark Zuckerberg prioritised growth and engagement over child safety on Facebook and Instagram. Reuters reports the testimony came in the landmark trial brought by state attorneys-general over alleged harms to children.
āāāā
šļø Government, procurement & public sector tech
US lawmakers are increasingly submitting AI-drafted bills to the House Office of Legislative Counsel, whose lawyers say they are spending more time repairing flawed text than drafting legislation from scratch. Politico reports the drafts can contain legal errors and unintended consequences.
āāāā
š New from ASPI
New research from ASPI, in brief.
Digital infrastructure in South Asia: Chinaās role and the potential for AustraliaāIndia cooperation ā Shruti Jargad, ASPI visiting fellow, maps Chinaās role across the regionās digital ecosystem and proposes AustraliaāIndia cooperation built around diversification, local ownership and institutional resilience. The recommendations give policymakers practical routes to expand credible infrastructure, financing and governance choices without demanding technological decoupling; Jargadās fellowship is supported by DFAT and the Centre for Australia-India Relations.
āāāā
š From The Strategist
This weekās cyber and tech analysis from ASPIās The Strategist.
Worth a try: the US will authorise some companies for cyber counterattacks ā Jason Healey argues the US experiment in authorising selected companies to counterattack cybercriminal groups is worth testing with guardrails.
How cheap, simple building materials can protect strike drones from lasers ā Tim Law argues low-cost fire-resistant building materials can harden strike drones against expensive laser defences.
Australia is not ready for an economic blockade of Taiwan ā Nathan Attrill argues Australia needs contingency planning for the economic consequences of a Chinese blockade of Taiwan.
āāāā
š§ On Stop the World
Related listening from ASPIās Stop the World podcast.
Anton Leicht on Australiaās AI middle power pathway ā Anton Leicht makes the case for Australia building data centres that can train and run AI, the national-capability question sitting behind NSWās new planning and resource rules for those facilities.
š Listen on Apple, Spotify, Amazon Music, RSS, or wherever you get your podcasts ā or watch on YouTube
āāāā
Earlier this week
These stories appeared in Tuesdayās Early Edition. Theyāre included here so Fridayās free Digest remains the complete record of the week. Early Edition readers can stop here.
What Weāre Watching
A weekly scan of notable developments weāre tracking across technology, policy and geopolitics.
š Global policy
š¦šŗ Australia
Guardian Australia found six erroneous or untraceable references in the emerging-technologies chapter of the A$3.48 million age-assurance trial used to support Australiaās under-16 social-media ban. The contractor, Age Check Certification Scheme, initially denied using AI, then acknowledged using ChatGPT to tighten prose after links carrying ChatGPT metadata were identified; it denies AI generated the research. The department is examining the concerns, while Senate evidence says officials have not independently verified the contractorās explanation, Guardian Australia reported.
The Telecommunications Industry Ombudsman has called for individual remedies when regional, rural and remote Australians face significant delays accessing mobile services, and for service standards that give consumers remedies when universal-service obligations are not met. The report says consumers outside major cities have fewer connectivity options and often wait longer for faults to be resolved, according to The Conversation.
Steiner Education Australia will not build a generative-AI learning tool, citing risks to healthy cognitive development, as government systems in New South Wales, South Australia and Queensland expand classroom AI. Queenslandās Corella tool uses a Socratic model, while educators warned that excessive cognitive āoffloadingā could weaken independent problem-solving. Other schools are pursuing controlled assistants for lesson preparation, explanation and early identification of struggling students, ABC News reported.
šØš³ China
Chinaās National Data Administration plans to make the country a leading supplier of high-quality training data by 2028 and share datasets internationally, including the multilingual WanJuan collection aligned with āmainstream Chinese valuesā. Researchers have already found that Chinese-language answers from Western chatbots reflect state-media narratives more strongly than their English answers. āThe downside of this will be that it gives greater power for authoritarian states to dictate a chatbotās values,ā Alex Colville, who researches Chinaās AI ecosystem at the Australian Strategic Policy Institute, told The New York Times.
World Liberty Financial is collaborating with Hong Kong-based WorldClaw, which offers 90 AI models and accepts World Libertyās USD1 stablecoin. Reuters found 43 models come from Chinese developers including Alibaba, Baidu, Z.ai, DeepSeek and Moonshot; several face US military-alignment, Entity List or technology-theft allegations. The Trump family owns 38 per cent of World Liberty and earns revenue from its tokens, while the presidentās sons have promoted WorldClaw, Reuters reported.
š«š· France
Franceās Constitutional Council has struck down legislation that would have barred children under 15 from social media, finding the restrictions disproportionate to freedom of expression and privacy. President Emmanuel Macron asked the prime minister to prepare a revised proposal, Le Monde reported.
šŗšø United States
Nevada has granted Tesla a robotaxi permit but capped its initial fleet at 10 vehicles rather than the 5,000 the company sought. The permit also excludes airport pickups and imposes operating restrictions, Axios reported.
US Commerce Secretary Howard Lutnick said the administration opposes Apple sourcing memory from sanctioned Chinese producers CXMT and YMTC, while Appleās chief operating officer reportedly left open the possibility of using commodity Chinese memory, The Wall Street Journal reported.
š§ AI models, agents & compute
Anthropic has raised its broad estimate of misalignment risk in high-stakes situations from āvery lowā to ālowā, citing recent cybersecurity incidents. Its latest risk report also says the company does not plan to release a stronger internal system called āModel 2ā, Axios reported.
Anthropic says future Claude models will probabilistically alter token choices to embed a detectable watermark in generated text longer than 200 tokens. It is applying the system globally because it says it cannot yet scope the EU transparency requirement by region. The approach has prompted criticism that provenance marking may affect wording while remaining easy for determined users to remove through rewriting, Daring Fireball reported.
š” Cyber posture
Franceās Finance Ministry has confirmed attackers stole detailed tax records for more than 600,000 taxpayers and businesses after accessing an internal DGFiP tool. Le Monde reviewed a sample containing civil-status details, addresses, phone numbers, taxable income, tax rates and dependants; the ministry said investigations are continuing with ANSSI and it will notify CNIL.
Greg Brockman says the OpenAIāHugging Face incident showed OpenAI underestimated its modelsā real-world cyber capabilities. OpenAI is responding by strengthening safety requirements, using models for code security and infrastructure defence, probing attack paths continuously and limiting advanced cyber capabilities to trusted defenders. Brockman argues organisations should deploy bounded defensive agents now while retaining human control over consequential actions, The Defenderās Window says.
šµļø Surveillance states
A website impersonating Brisbane consultancy Horizzen was among 13 domains seized by the US Department of Justice in an investigation into alleged Chinese intelligence recruitment. The FBI says the sites used aliases, stolen identities and AI-generated photographs to solicit defence, trade and Indo-Pacific expertise, sometimes moving recruits toward paid requests for sensitive information. Five Eyes agencies have separately warned that Chinese intelligence officers pose as recruiters and consultants through convincing cover companies and professional platforms, Guardian Australia reported.
Documents reviewed by Reuters show Binance gave Russian investigators transaction records and identity documents used to bring terrorism-financing charges against a Russian IT specialist who sent more than US$700 to Ukrainian military fundraisers, including the Azov Regiment, which Moscow designates a terrorist organisation. The material included his passport, Bulgarian residency permit, address and telephone number, despite Binance saying in 2023 that it had fully exited Russia. A lawyer said EU privacy rules may have restricted the transfer if the customer was registered as an EU resident; Binance disputes that interpretation, Reuters reported.
āļø Platform accountability
Apple has proposed commissions of 5 to 15 per cent on purchases made outside its US App Store, after an appeals court said it could recover reasonable costs. In a separate Epic case, a judge ordered Google to remove āanticompetitive frictionā from installing rival app stores and make changes within a week, according to 9to5Mac and The Verge.
Apple will change its App Tracking Transparency prompts after Germanyās Federal Cartel Office found its own apps received more favourable treatment than third-party developers. The commitments require neutral language and design across almost all EU countries, take effect within four months and will be monitored for seven years. France and Italy have separately fined Apple over the framework, Reuters reported.
A coalition of 29 US attorneys-general has begun trial against Meta over alleged addictive design and childrenās privacy violations. The states seek nationwide remedies that could include deleting data and models trained on information from children under 13 and removing features such as infinite scroll, autoplay and engagement-optimised algorithms. Meta says the claims are unsupported and the financial demands disproportionate, CNBC reported.
Thatās all for this week. For more timely analysis and commentary, check out The Strategist and ASPIās Stop the World podcastāor our other Substack newsletters:
The Cyber & Tech Digest is brought to you by the Cyber, Technology & Security Programs team at ASPI and supported by partners.
