Welcome to the latest edition of ASPI’s Cyber & Tech Digest.
Each week, ASPI curates and contextualises the most important developments in cyber, technology, and geopolitics — highlighting what matters and why.
This edition covers the period: 20 June 2026 to 26 June 2026.
Follow the Australian Strategic Policy Institute on Bluesky, LinkedIn, and X.
The AI the world will actually run on
Remember when, only a few years ago, money was basically free?
Mortgage rates started with a 1, every app came with a sign-up credit, Uber felt absurdly cheap, and streaming services cost less than a sandwich. Debt was cheap, capital was abundant, and many of the services consumers got hooked on were subsidised. Investors were paying companies to buy our attention and our habits.
That is more or less where we are now with AI.
The experimental AI workflows many of us are maxing out are being subsidised by sovereign-wealth-fund and venture-capital backing. One estimate from SemiAnalysis calculated that a US$200 monthly AI plan can buy the equivalent of about US$8,000 worth of Claude Code tokens and US$14,000 worth of Codex tokens. It is akin to, as Cory Doctorow put it, selling US$100 bills for a dollar apiece. Put that way, it is no surprise demand is so high.
But even at these bargain-basement prices, token costs are proving too much for many businesses. Uber reportedly burned through its 2026 Claude Code budget only a few months into the year — a ‘head-exploding moment,’ according to its COO, because all that token usage was not clearly translating into useful features.
Companies have quickly realised that usage does not automatically translate into return on investment. The AI firms know it too, and are scrambling to make the economics work: OpenAI is building an advertising business, while Anthropic is inching from generous flat-fee subscriptions towards usage-based billing. ‘Intelligence,’ says OpenAI’s Sam Altman, is something customers will buy ‘on a meter’.
So users and companies are scrutinising their workflows to work out what can be done cheaply and what must be reserved for higher-end models. Sophisticated work might still go to a frontier model such as Fable 5, which — before it was disabled under a US export-control directive — cost US$10 per million input tokens. The routine grind can be pushed to an open-weight model such as DeepSeek’s low-cost tier, at US$0.14 per million input tokens. Coinbase’s Brian Armstrong has predicted that, within 12 to 18 months, 80 per cent of AI workloads will run on models that are 99 per cent cheaper than the latest frontier systems.
This is the next phase of AI adoption: not everyone using the most powerful model for everything, but companies routing each task according to cost, capability and risk, and, increasingly, according to whose model they are willing to build on. Frontier models will still matter for research, complex agents and high-stakes tasks where being wrong costs more than the token bill. But the daily grind of corporate AI will be pushed down to cheaper models that are good enough.
And this is where the economics become geopolitical.
Open-weight models are no longer a hobbyist alternative. They are becoming part of the enterprise stack — cheaper, easier to customise, easier to run locally, and harder to switch off from afar. Increasingly, the most attractive options are Chinese. Frontier labs do the expensive work of discovering new capabilities; once demonstrated, those capabilities can be distilled, compressed and commoditised by cheaper models. Chinese companies have proved especially good at this game. DeepSeek, Alibaba’s Qwen and Moonshot’s Kimi are being downloaded and routed into global workflows because they are cheap and they work.
But a default layer is never neutral. A model carries the values, blind spots and controls of whoever trained it, and Chinese models are trained to comply with China’s information rules. Ask them about Tiananmen, Xinjiang or Taiwan and they deflect, sanitise or, more insidiously, fabricate: returning Beijing’s preferred answer with confidence and no sign that anything has been suppressed. That matters little when one hobbyist runs one query. It matters enormously when a cheap Chinese model becomes the default that millions of businesses, apps and public services quietly route through. Those defaults then propagate, not as overt propaganda but as the quiet baseline of what the world’s AI tools treat as true. And the risks are not only about speech. A weights file can be tuned to send data home, or to carry flaws that surface only once it is running deep inside someone else’s systems. A proprietary model behind an American API can at least be audited and, if need be, switched off. A model downloaded and embedded everywhere cannot.
The infrastructure bill for all this is heading towards the trillion-dollar scale. Data centres, chips, power and cooling are being built on the assumption that demand will keep rising fast enough to justify the spend. To become genuinely profitable, AI companies must earn revenues that exceed both their model-development costs and this vast build-out.
That is why the contest cannot stop at who has the best frontier model this month. What matters more is who supplies the default layer of cheap, capable, widely available AI that businesses, developers and governments build on every day.
Chinese open-weight models may grow more attractive for another reason too: the US government has shown it can force access to a frontier model to be suspended. That may be defensible in some national-security cases. But for users outside the United States, it is a reminder that depending on proprietary American models carries a geopolitical risk of its own. Open weights, once downloaded and shared, are far harder to switch off than a model sitting behind a US company’s API.
Washington has noticed the problem. The White House has said the United States should foster a supportive environment for open models, and OpenAI has released open-weight models of its own. But the response remains underpowered.
Liberal democracies need something closer to a public-option AI stack with cheap, capable, open-weight models built and maintained by institutions whose mission is public value rather than private return. That takes real money and real will. Public-interest labs that can train on high-quality data lawfully, run their own compute, and put trusted models into the hands of developers and governments around the world.
The point is not to replace the frontier labs. The world will still need the most capable models liberal democracies can build. But most AI use will happen in the cheaper, messier layer underneath — the layer businesses route to by default, startups build on, governments can afford, and users in poorer countries can actually reach.
If liberal democracies try only to win the race at the top, they may still lose the market underneath. Releasing a handful of open models and calling it a strategy, as Washington has so far, treats that layer as a box to tick when it should be a category to win. The democratic world has the talent, the data and the capital to make the cheap, trusted default its own. But the window is closing. Win the frontier while the world’s everyday AI runs on Chinese models, and the lead won’t be worth much.
— Fergus Ryan CTS
What We’re Tracking
‘The timeline is not years, it is months’: the Five Eyes’ AI-cyber warning
What happened: In a rare joint statement, the cyber-security chiefs of all five Five Eyes agencies — Australia’s ACSC, the US NSA, the UK’s NCSC and their Canadian and New Zealand counterparts — warned that frontier AI models are about to ‘fundamentally transform both offensive and defensive cyber capabilities’, and that ‘the timeline is not years, it is months’. AI is shrinking the window between a vulnerability being discovered and exploited, they said, urging leaders to act now and to ‘use AI deliberately to strengthen defence’. The intervention followed Washington’s decision to bar foreign nationals from Anthropic’s Fable 5.
The warning lands as the China–US capability gap looks smaller than it has in over a year. Zhipu’s new open-weight model, GLM 5.2, was judged by the research firm Artificial Analysis the leading open-weights model on the market, and it was released a day after the Fable ban. By the Economist’s account, drawing on analysis from the Norwegian Defence Research Establishment, Chinese models sit four to six months behind the US frontier on public benchmarks and eight to ten on private ones. It’s a gap that’s not widening as many had expected.
China is now moving to field its own version of the very capability that worries Washington. At a Beijing conference, 360 Security Technology founder Zhou Hongyi unveiled ‘Tulongfeng’ — which he called ‘China’s version of Mythos’, Anthropic’s vulnerability-finding system — alongside a defensive tool, ‘Yitianzhen’. He said Tulongfeng had already found 3,432 vulnerabilities, 105 of them confirmed by Chinese authorities, though Reuters could not verify the claims, and framed the effort as a response to the risk of ‘one-way transparency’ if US models could scan systems while Chinese firms were denied the same.
Why we’re tracking this: Put the three together and the picture is clear: the people who defend Western networks say the threat clock is down to months, not years; the capability gap with China is no longer widening; and Beijing is openly racing to build the same offensive-cyber AI. Yet the current export settings cut the other way. Washington’s directive appears to have cut Australian users and organisations out of ordinary access to Fable 5 and Mythos 5 — absent exemptions or classified arrangements that have not been publicly disclosed — at the very moment the Five Eyes are urging defenders to use AI to strengthen their networks. (Anthropic has since restored Fable 5 behind nationality-based controls; Mythos access for foreign nationals remains restricted or unclear.)
What people are saying:
‘The timeline is not years, it is months.’ — Five Eyes cyber-security agencies, joint statement
‘I think we have to anticipate that the next Mythos or the next Fable is just around the corner.’ — Olivia Shen, United States Studies Centre, University of Sydney
‘This kind of powerful weapon that can change the landscape of cyber offence and defence cannot be held only by others.’ — Zhou Hongyi, founder of 360 Security Technology
My view: The Five Eyes agencies do not issue joint statements lightly, and the message here is unusually blunt: the window is months, not years, and defenders need to be using the best AI now. Read alongside GLM 5.2’s rise and 360’s announcement, it describes a threat that is both imminent and increasingly within China’s reach.
That makes the current export settings look backwards. Keeping Mythos-class capability confined to US-controlled channels may dampen some proliferation risk, but a blunt foreign-national rule also leaves trusted allies slower to defend against the very capabilities China is racing to replicate. As I told the ABC earlier this month, if Washington now sees frontier AI models as strategic capabilities, it needs an allied-access regime for Five Eyes partners, not a one-size-fits-all foreign exclusion rule. The Five Eyes’ own warning only sharpens that case. Harden the alliance’s defences first, not last.
— Fergus Ryan CTS
What We’re Watching
A weekly scan of notable developments we’re tracking across technology, policy, and geopolitics.
⸻
🚀 Strategic competition
The Netherlands agreed to join Pax Silica, a US-led group coordinating allied AI supply chains. Trade Minister Sjoerd Sjoerdsma raised concerns about the proposed Match Act, telling Commerce Secretary Howard Lutnick and lawmakers that the Netherlands shared US goals on sensitive technology but opposed forcing allies to adopt tighter China chip-export controls. The US separately proposed that the European Union join an AI partnership focused on critical minerals, energy, computing and semiconductor manufacturing.
ASML denied that it had shipped an extreme ultraviolet (EUV) lithography machine or related component to China after Lutnick reportedly questioned the company about whether advanced equipment had reached the country. ASML also said the claims were inaccurate and damaging and that it knew the location of every EUV tool it had built. The Dutch foreign ministry said covered semiconductor-equipment exports required EU or national licences.
China’s Commerce Ministry added 10 US defence firms to its control list, while the Finance Ministry excluded 46 US companies from government procurement after the Department of Defense expanded its list of alleged Chinese military companies. The measures covered rare earths, drones, robotics and aerospace. Alibaba, meanwhile, sued the US government over its inclusion on the list, denying that its products were built for weapons, defence or intelligence and calling the designation arbitrary and reputationally damaging.
President Donald Trump said his administration had helped bring Nvidia investment into Intel and that Elon Musk had agreed to build TerraFab with Intel’s technology team. Separately, the AI data-centre boom lifted Asian semiconductor and hardware suppliers, including SK Hynix, Samsung, Taiwan-linked suppliers and Micron, while The New York Times reported that China remained largely outside the most lucrative Nvidia-centred advanced AI server supply chain.
SpaceX warned European officials that an EU satellite-spectrum proposal could reduce direct-to-device services and create interference risks, including for emergency communications in Ukraine. The proposal would reserve part of a band for European operators and limit frequencies available to US and Chinese players. SpaceX said the plan prioritised incorporation location over technical and regulatory considerations.
Russia expanded domestic AI training through a presidential commission, curriculum changes and a Moscow State University AI faculty linked to Vladimir Putin’s daughter, Katerina Tikhonova. TIME reported that sanctions, weak domestic hardware capacity and reliance on grey-market US chips or possible Chinese supply continued to constrain access to advanced chips.
Europe 2031, a speculative scenario, circulated widely in European policy circles after the US government blocked foreign nationals from using Anthropic’s Fable model. Its authors argued that Europe risked economic and strategic decline without more AI infrastructure and technological sovereignty.
⸻
🧠 AI models, agents & compute
Anthropic accused Alibaba and its Qwen AI unit of creating almost 25,000 fake accounts to access Claude, targeting capabilities including agentic reasoning, software engineering and long-horizon tasks. The company said the campaign generated more than 28.8 million exchanges between late April and early June and could accelerate China’s access to Mythos Preview capabilities. Anthropic called for tighter US chip controls and penalties for AI laboratories found engaging in illicit distillation.
360 Security Technology said it had developed two AI security tools under the Yitian Tulong banner as a domestic answer to Anthropic’s Mythos. Founder Zhou Hongyi said Tulongfeng was designed to find software vulnerabilities and Yitianzhen to automate cyber defence and incident response. The company said Tulongfeng had found 3,432 vulnerabilities, including 105 confirmed by Chinese authorities, though Reuters could not independently verify the claims.
Some early Project Glasswing users retained access to Mythos Preview after the government order led Anthropic to shut down other Mythos and Fable versions. Dragos and Cisco confirmed continued access, while people familiar with the matter said banks and technology firms were also using the preview for vulnerability work. Bloomberg reported that the order did not explicitly address the Preview version.
The Trump administration held talks with Anthropic about re-releasing Claude Fable 5, including what proof the company would need to provide on jailbreak risks. Cofounder Tom Brown and public policy chief Sarah Heck became the company’s main interlocutors in place of chief executive Dario Amodei, according to WIRED. The White House and Anthropic also negotiated common benchmarks for assessing AI jailbreaks, focused on how safeguards were bypassed, what capabilities were exposed and the consequences of a breach. President Trump separately said relations had improved after discussions with Amodei, after earlier describing Anthropic as a possible national security threat.
⸻
🛡 Cyber posture
Microsoft’s Digital Crimes Unit used Copilot to connect the malware behind Amadey and StealC, allowing the company to file one RICO lawsuit targeting both tools as part of an alleged conspiracy. Microsoft said the tools were linked to more than 140,000 infected computers worldwide during the first two weeks of last month. The operation involved Europol, German, Dutch and Danish police, IBM and Proofpoint.
Icarus claimed responsibility for a breach at Klue that exposed data from corporate customers including Gong, Jamf, HackerOne, Insurity, OneTrust, Recorded Future, Snyk, Sprout Social and Tanium. Klue said attackers entered its systems earlier this month using a compromised legacy credential tied to an integration tool connecting customers’ cloud data to Klue accounts.
Romania’s national cyber-security centre ordered more than 100 hospitals to disconnect from the internet during a ransomware attack in early 2024 that spread through widely used medical software. Investigators found 26 hospitals infected with BackMyData, while staff used paper records, Excel and offline workarounds. Authorities declined to pay a €160,000 bitcoin ransom, and most hospitals returned online within five days with no reported deaths or serious patient harm.
Brazil’s government took the National Civil Defense warning platform offline after hackers were suspected of sending an unauthorised extreme alert containing ‘misantropi4’ to phones in Paraná, São Paulo and Rio de Janeiro. Authorities said there was no emergency and were investigating the message’s origin. The government is working to restore the platform once security conditions are re-established.
⸻
🕸️ Illicit networks & infrastructure
CoinEx handled more than $3.84 billion in transfers from wallets with identifiable links to Iran since 2019, according to TRM Labs data reviewed by The Wall Street Journal. The flows included funds tied to Iran’s Central Bank, Nobitex and accounts later attributed by US officials to the Revolutionary Guard. Founder Haipo Yang said the exchange was widely used by Iranians but denied any relationship with Iran’s government.
Officials and researchers said residential proxy networks were being used by state-backed hackers from Russia, China, Iran and North Korea to obscure attribution and target US networks.
⸻
⚖️ Platform accountability
Meta left potentially sensitive data collected from employee laptops accessible internally to anyone at the company. The data, gathered through the Model Capability Initiative, was believed to include keystrokes, mouse clicks and screen content from US employees. Meta said it was investigating and had no indication the data was improperly accessed.
Attorney General for England and Wales Richard Hermer told his office to stop posting on X except to counter disinformation. The Guardian reported that the decision followed disorder in Southampton and Belfast and concerns that bad actors were using X to divide communities. Hermer’s office became the first UK government department to stop using the platform amid wider government concern about social media and inflammatory content.
⸻
🧑⚖️ Courts, enforcement & regulation
The US Supreme Court ended a lawsuit by Falun Gong members accusing Cisco Systems of helping China surveil and persecute the group through the Golden Shield system. In a 6–3 ruling, the court held that the Alien Tort Statute did not create aiding-and-abetting liability for Cisco’s disputed conduct. It also ruled 8–1 that the Torture Victim Protection Act did not allow claims against two Cisco executives to proceed.
⸻
🏛️ Government, procurement & public sector tech
President Donald Trump signed two executive orders aimed at accelerating US quantum-computing development and preparing for quantum-enabled hacking. One directs agencies to work with companies and academics to deploy a quantum computer for scientific research by 2028; the other directs agencies to harden government and private-sector security systems against future quantum attacks.
The Federal Energy Regulatory Commission approved orders to speed data-centre connections to US power grids while seeking to limit pressure on utility bills. The target is to process power requests within 90 days, compared with a process that can currently take years.
UK officials were examining whether and how Palantir could be removed from NHS systems amid questions over its £330 million NHS Federated Data Platform contract. Critics said the tender favoured Palantir and gave the company too much control over NHS data infrastructure, while NHS England and Palantir said the platform had improved patient outcomes. The questions also covered performance, data sovereignty and reliance on US technology.
Western Australia Police will trial real-time facial-recognition cameras in public spaces, making Western Australia the first Australian jurisdiction to deploy the technology. The AI-enabled system will compare faces captured by cameras mounted on or near marked police vehicles against a database. When the system detects a match, the image will be retained and nearby officers alerted to confirm the person’s identity.
⸻
👷 AI, work & skills
China accelerated deployment of AI-enabled industrial and humanoid robots across manufacturing, logistics, retail and healthcare, with Beijing directing local governments and state-owned enterprises to incorporate embodied AI. Companies including Sany, Huazhu and JD.com were automating factory, hotel and delivery work, while analysts and academics warned of job displacement, youth unemployment and labour-income pressures. JD.com founder and chair Richard Liu separately said robots would eventually replace the company’s 700,000 delivery workers and that the company had signed contracts with about 120 schools to retrain couriers for robot repair and maintenance roles.
OpenAI, Anthropic, Amazon, Microsoft and other employers joined Raise Us, a new nonprofit led by Gina Raimondo that had raised $500 million to address potential AI-driven workforce disruption. The group plans to work mainly with state governors, community colleges and employers on pilots including service-year expansion, wage insurance and company retraining, alongside a policy lab.
Microsoft chief executive Satya Nadella argued for cheaper models, more user control and broader public trust and criticised an AI race dominated by frontier model builders. Microsoft has released low-cost models and Copilot Cowork and was considering hosting a version of DeepSeek. Nadella said AI companies should focus on reorganising jobs rather than using the technology mainly for job cuts.
⸻
💰 Tech business & markets
Meta assigned a small team to develop an experimental prediction-markets app called Arena. The smartphone app would operate separately from Facebook, Instagram, WhatsApp and Messenger and would likely use points rather than money at launch. Meta declined to comment, and employees said the app might not be released.
Early Chinese backers of Manus were planning to buy the company back from Meta for the $2 billion acquisition price after a Chinese government order to reverse the deal. The Information reported that Manus’s annualised revenue run rate had risen to between $400 million and $500 million, from $100 million when Meta acquired it late last year. Beijing objected partly because Manus had not notified regulators before the sale and because officials feared a precedent for AI start-ups selling technology and talent to the US.
DeepSeek reportedly closed its first external funding round at a valuation above $50 billion. Founder Liang Wenfeng told prospective investors that participation required them not to poach staff or encourage employees to start companies, according to CNBC, citing 36Kr. The condition followed the loss of key researchers to rivals as Chinese technology companies competed for advanced AI talent.
Former White House senior policy adviser Dean Ball joined OpenAI to lead a new Strategic Futures team focused on frontier AI policy and internal governance. OpenAI chief strategy officer Jason Kwon said Ball would help test the company’s thinking on risk, governance and frontier policy.
The Pew Research Center found that 49 per cent of US adults reported using AI chatbots, up from 33 per cent in 2024. ChatGPT was the most used at 44 per cent, followed by Gemini, Copilot and Meta AI. The survey also found that 63 per cent said AI was advancing too quickly and 71 per cent said increased AI use would make personal information less secure.
⸻
🧒 Online harms & child safety
Norway will impose a near-ban on generative AI tools for primary-school pupils and restrict use by older children. Prime Minister Jonas Gahr Stoere said young children risked skipping important learning steps. Pupils aged six to 13 should generally not use AI, lower-secondary pupils may use it under teacher supervision, and upper-secondary students should learn appropriate use.
Companies marketed AI humanisers, autotypers and writing tools to students that could make AI-assisted cheating harder to detect. The New York Times reported that some tools could rewrite AI-generated text, simulate human typing patterns and evade detectors. It also reported that some companies selling academic-integrity or writing products offered features that could generate, humanise or rewrite student work.
⸻
🌏 Global policy
🇦🇺 Australia
Independent senator David Pocock alleged that a whistleblower had described competing federal cabinet proposals on AI copyright, including a text-and-data-mining carve-out linked to data-centre investment and an expanded licensing model. Industry and Science Minister Tim Ayres called the claim reckless speculation and said the government would not undermine copyright protections; the ABC said it had not verified the whistleblower, the submissions or the claims. Industry and attorney-general officials were preparing cabinet submissions on different approaches, according to Guardian Australia, while Ayres’s spokesperson said the government had ruled out a text-and-data-mining exception.
🇺🇸 United States
The Trump administration used export controls, voluntary testing frameworks, procurement guidelines and executive actions in case-by-case interventions affecting the AI industry, according to Axios. The article identified uncertainty around Anthropic’s models, federal procurement standards and state AI laws as areas being shaped outside formal rulemaking.
Influence over administration AI policy shifted as David Sacks stepped back from daily involvement and Sriram Krishnan prepared to leave. Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent, White House chief of staff Susie Wiles, National Economic Council official Ryan Baasch and National Cyber Director Sean Cairncross were described as key figures, with the Anthropic dispute elevating Lutnick and exposing internal tensions over AI security policy.
Vice President J. D. Vance said he had not seen evidence of widespread AI-driven job destruction but was concerned that AI’s benefits could flow disproportionately to the wealthy. The Atlantic reported that Vance’s approach combined scepticism of regulation, concern for workers, national-security priorities and wariness of Big Tech concentration, while Republican divisions over state and federal AI rules remained unresolved.
A forthcoming book by Maggie Haberman and Jonathan Swan described efforts by Mark Zuckerberg and Jeff Bezos to ingratiate themselves with Donald Trump after his election win. The book said Trump mocked those efforts privately and that Bezos criticised The Washington Post to Trump during a Mar-a-Lago dinner.
That’s all for this week. For more timely analysis and commentary, check out The Strategist and ASPI’s Stop the World podcast—or 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.
