AI Policy and Governance Newsletter — June 2026

June 2026 Newsletter

16 June 2026

For barely a week, Australian banks, hospitals and infrastructure operators had access to the most capable AI models Anthropic has built. Then, on 12 June, the US government ordered Anthropic to cut off every foreign national, so Claude Mythos 5 and Fable 5 vanished worldwide.

The order landed days after Anthropic called for the option of a coordinated global pause on frontier development, and its chief executive Dario Amodei asked Washington for FAA-style testing to block unsafe models before release.

Australia's AI Safety Institute is now operational with Dr Kate Conroy, and Senator David Pocock is pressing the Government on whether the Institute's $29.6 million over four years is sufficient.

Assistant Minister Charlton is putting trust at the centre of the Government's data-centre and AI buildout strategy, including by highlighting that Australia can take a different path to the US to avoid the backlash that their data-centre buildout has caused. Meanwhile, the Senate has launched an inquiry into data centres and the Government's MoUs with their operators. This inquiry comes as Atlassian co-founder Scott Farquhar warns that Australian copyright law makes the country "close to impossible" for AI developers.

Pope Leo XIV's 25 May encyclical Magnifica Humanitas asks who AI is being developed for, and at whose expense, calling for "robust legal frameworks, independent oversight, informed users and a political system that does not abdicate its responsibility". A week later Donald Trump signed an executive order creating a voluntary federal framework for AI model review before public release.

—

Welcome to the AI Policy and Governance newsletter from Good Ancestors. We track the biggest developments in AI policy and safety, at home and abroad.

—

Featured Australian publications

• System Risk Outlook (Australian Prudential Regulation Authority, 21 May): APRA names rapid AI adoption as a primary source of systemic vulnerability for regulated entities, with Chair John Lonsdale flagging that "rapid developments in AI are outpacing the ability of many entities to manage the risks".
• Policy Advisory 001-2026: Cyber Security Readiness in the Frontier AI Era (Department of Home Affairs Protective Security Policy Framework, 26 May): Mandatory whole-of-government advisory warning that frontier AI is collapsing the window from vulnerability discovery to exploitation from days to hours, creating a "vulnerability storm".
• GPT-5.5 Saturates Our Offensive Cybersecurity Time Horizons (Lyptus Research, 27 May): Lyptus updated their previous offensive cyber time-horizon study with GPT-5.5 data, finding that the model achieves a 50% time horizon of 5.1 hours at standard 2M-token evaluation budgets and that performance continues to scale with tokens. The authors argue frontier cyber capability is now outpacing what offensive-task evaluation benchmarks can measure.
• Australian Community Attitudes to Privacy Survey 2026 (Office of the Australian Information Commissioner, 28 May): Just 4% of Australians believe AI companies are worthy of their trust (the lowest of any sector measured) and 87% have become more concerned about privacy in the last five years. Privacy complaints to the OAIC are up 73% this financial year.
• Quarterly Essay 102: The God We Made — The Threat and Promise of Artificial Intelligence (Anna Goldsworthy, 1 June): A long-form essay tracing AI through philosophical, cultural and political lenses.
• Expanding AI Sovereignty to AI Agency (Tech Policy Design Institute, 15 June): The first independent scorecard of Australia's AI capabilities, mapped against the Government's National AI Plan. Australia rates strong on critical minerals, data and norm-shaping, weak only on AI-chip manufacturing, with regulatory oversight flagged as a key gap.

—

News & commentary

Washington pulls access to Anthropic's frontier models shortly after Australia gains access

On 12 June, the US ordered Anthropic to suspend all access to its two most capable Claude models (Mythos 5 and its heavily guardrailed version Fable 5) for any foreign national, whether inside or outside the United States (including Anthropic's own foreign-national staff). The export-control directive, signed by Commerce Secretary Howard Lutnick, cited national-security authorities but gave no detailed explanation. Because Anthropic cannot reliably tell foreign nationals from US persons across its user base in real time, it disabled both models worldwide.

Australia had access to these models for barely a week. On 2 June Anthropic had expanded Project Glasswing to Australia (along with about 150 organisations across more than 15 countries) which granted access to Claude Mythos Preview to critical-infrastructure operators in power, water, healthcare and communications. Treasury, the Reserve Bank of Australia and the stewards of Australia's Systems of National Significance had been briefed in the preceding weeks, after months of engagement by Anthropic's chief legal counsel and former US ambassador to Australia, Jeffrey Bleich.

A week later on 9 June, Fable 5 (the Mythos-class Claude model fitted with classifiers that route cyber, biological and chemical requests to the less-capable Opus 4.8) was released worldwide to the general public and three days later it was gone.

The directive named no specifics, but the US Government appears to have acted on reports that Fable 5's safeguards had been defeated. UK AI Safety Institute head of safeguards Xander Davies broke them within hours of the public release, despite Anthropic's claim that an external bug bounty had found "no universal jailbreaks in over 1000 hours" of testing. Anthropic complied but disputed the order, warning that pulling a model over a post-release jailbreak would, as a standard, stop new frontier models being deployed at all. In Australia, the AFR reported business leaders and ministers likening the loss of access to a strategic supply blockade.

Comment:

The International AI Safety Report 2026 warns of information asymmetries – specifically that the knowledge gap between AI company insiders, AI experts and the general public is growing. In most tabletop exercises Good Ancestors has run exploring AI futures, frontier models are withdrawn from the public. This leaves the public and smaller countries largely blind to AI capabilities and risks.

At a minimum, the ongoing fight over Mythos highlights the urgent need for transparency requirements and for public and independent evaluation of AI capabilities, even when they're not available on the open market. While the dust is still settling, it's fair to ask what this means for sovereign AI, including the need for countries to develop "fast chaser" AI models they can control, and/or to deepen ties with existing frontier AI companies and diplomatic engagement with the US on model access.

Anthropic asks to be regulated and floats the option of a coordinated pause

Anthropic used its new Anthropic Institute on 4 June to call for the option of a coordinated, verifiable slowdown of frontier AI, saying it would suspend work on more capable systems if it could be confident rivals would do the same: "If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing." The argument turns on recursive self-improvement (the point at which models can design and build their own successors with little human input). To make the case, Anthropic's Jack Clark and Marina Favaro published internal data: more than 80% of the code merged into Anthropic's codebase is now written by Claude, and in April the model resolved more than 800 software bugs the company estimated would have taken a human engineer four years.

The Albanese Government welcomed the intervention. Assistant Minister for Science and Technology Andrew Charlton, who leads Labor's AI strategy, said the pace of progress "must not outstrip understanding or undermine safety across society, and no company should develop AI that is unsafe", adding that "the Government welcomes calls for a more considered approach to frontier AI development".

Anthropic's CEO Dario Amodei went on to argue in an essay on his blog that the era of transparency-only rules was over and called for FAA-style mandatory testing of frontier models (across cybersecurity, bioweapons, loss of control and automated R&D) with a model's release "blocked or reversed" if it fails to meet the standard. Anthropic pledged US$350 million to the effort and published a draft legislative framework. OpenAI also endorsed coordinated action "including slowing frontier development when needed", but argues that "democratic governments — not private companies acting alone — must ultimately determine the rules".

Trump's former AI tsar David Sacks dismissed the warning as a play for a bailout — a sign "you might be trying to get your frontier AI lab nationalised" (noting Anthropic has filed for an IPO that could value it above US$1 trillion) and PauseAI Australia welcomed the acknowledgement that "a global pause button is required" but said the terms of any freeze should be negotiated by governments, not AI companies. Anthropic itself was candid about the difficulty: "Training runs are far easier to conceal than missile silos," and a slowdown that "simply lets the least cautious actors catch up" would leave everyone less safe.

Comment:

Seeing Charlton actively commenting on this news is a sign that Australia is following closely. But his statement should raise eyebrows. He says that AI development shouldn't outstrip safety efforts, but we know that it is. And he says that no company should develop AI that is unsafe, but we know that they are.

The national AI plan says that existing regulators will manage AI risks within their domain, but that government "will take decisive action to ensure safety and accountability as new technologies and frontier AI systems emerge" and that "if more regulation is needed to address bad actors or broader harms, the government will not hesitate to intervene".

Perhaps this is an early sign that Government is warming up for more action.

Australia's AI Safety Institute is officially launched — but is it adequately resourced?

Assistant Industry Minister Andrew Charlton announced that the Australian AI Safety Institute (AISI) is now operational (as of 2 June) with three stated goals: analyse and test new AI models and applications, support regulators in responding to emerging harms, and shape AI development and governance in Australia's interests. The Institute's inaugural general manager is Dr Kate Conroy, a philosopher who concurrently holds the role of Responsible AI Lead at the Royal Australian Air Force (RAAF). The launch follows the Australia–UK AI safety bilateral signed the week prior.

The same day, at Senate Estimates, independent ACT Senator David Pocock pressed Industry Minister Tim Ayres on the AISI's resourcing and independence. Pocock noted the Institute's $29.6 million over four years is well below the UK AI Security Institute's £66 million (about A$124 million) annually, questioned the decision to keep the AISI inside the Department of Industry, Science and Resources rather than establish it as an independent regulator, and raised Conroy's concurrent RAAF role. Pocock said the Government was "out of step with most Australians" on AI safety. Minister Ayres defended the appointment and the Institute as "well resourced" with "common features" to the UK and Canadian institutes.

Comment:

A well-designed AISI is one of Australia's strongest levers for impact on AI policy and safety. Getting it operational is a milestone that Good Ancestors has been advocating for since the Bletchley Park summit three years ago. Pocock is right that the funding falls short, Good Ancestors' AISI Expert Survey of 139 professionals found 53% of respondents recommended an annual budget exceeding $50 million. This is consistent with the UK AISI's budget, which equates to AUD 50–65 million per year when scaled to Australia's population or GDP.

But the AISI is only just getting started. A new institute deserves time to set its priorities, build the technical team, and produce its first round of work before being judged on substance and hopefully bringing in more funding as its value is proven. We are cautiously optimistic, and watching closely.

Charlton bets the data-centre buildout on trust

In a keynote to the Australian Financial Review's AI Summit on 2 June, Assistant Industry Minister Andrew Charlton placed trust at the centre of the Government's data-centre buildout strategy. He cited the Melbourne University and KPMG global survey showing just 30% of Australians believe AI's benefits outweigh the risks (the lowest of 47 countries). He pointed to mounting backlash in the US, where at least 11 states have moved to pause or ban new construction existing in at least 11 states, US$18 billion in projects blocked, and family power bills in America's largest electricity market projected to rise around US$70 a month by 2028.

In NSW's Lane Cove West (where data centres have clustered against homes, a school, childcare centres and bushland) the local council is considering a moratorium. Charlton said he could already see "the early warning signs" of what he called "the American story, beginning to write its first Australian chapter" and that Australia has a window to get ahead of it. He defended the Government's Expectations of Data Centres and AI Infrastructure Developers (where operators bring their own new renewable power, pay full transmission and distribution costs, engage genuinely with local communities, and do not complicate water access) as "regulation that builds trust and confidence".

At the same Summit, Atlassian co-founder and Tech Council chair Scott Farquhar said Australian copyright law is "close to impossible" for AI developers and that without reform AI training "data centres will go elsewhere".

In May the Senate's Environment and Communications References Committee launched a six-month inquiry into data centres' environmental, industrial and community impacts and the Government's MoUs with centre operators, reporting in November. Greenpeace called for an urgent moratorium on data-centre development, finding Sydney's proposed 1GW Mamre Road campus alone would generate peak annual grid emissions equivalent to 560,000 petrol cars. The NSW Net Zero Commission has proposed strict matching-renewables and peak-demand caps, Victoria now claims a $1.2 billion data-centre capex edge over NSW, and the ABS recorded the highest sectoral capex on record in information, media and telecommunications, up 66% quarter-on-quarter to $8.2 billion.

Comment:

Charlton is broadly correct in framing trust as a precondition for growth rather than a constraint on it, and asserting that Australia has the power to choose how a data-centre roll out goes. We argued in our submission to the NSW Legislative Council's data-centre inquiry that the key question for Australia is what conditions make the buildout serve the national interest. The energy and water conditions in the Government's Expectations look workable. Anthropic committed in February to cover 100% of grid upgrade costs and offset consumer power-price increases, and the largest US AI companies have since signed a parallel pledge. But Government still needs to navigate copyright and identify other terms that will put Australia in a good position to ensure this technology is broadly safe and beneficial.

A moratorium on data-centre development would misread the trade-off. Global chip supply is constrained, so data centres that Australia declines would get deployed elsewhere, in places with no guarantee of stronger safeguards and no Australian say over the terms.

Good Ancestors-commissioned YouGov polling released last month found 61% of Australians support enabling AI training under arrangements that compensate creators, and 57% support a library-style fund modelled on the Public Lending Right scheme.

Pope Leo XIV calls for the "disarming" of AI in landmark encyclical

Pope Leo XIV released Magnifica Humanitas on 25 May: a 40,000-word encyclical that calls for the "disarming" of AI, particularly autonomous weapons and mass surveillance. The document warns AI can be programmed to maximise profit alone, generating "'necessary sacrifices'... in pursuit of the supposed optimization of the species", and calls for "robust legal frameworks, independent oversight, informed users and a political system that does not abdicate its responsibility".

Anthropic co-founder Chris Olah was invited to speak at the Vatican launch. In his remarks, Olah conceded that "every frontier AI lab — including Anthropic — operates inside a set of incentives and constraints that can sometimes conflict with doing the right thing", and called for "informed critics who will tell the labs when we are failing... moral voices that the incentives cannot bend".

Transformer's Shakeel Hashim argued Magnifica Humanitas "feels woefully dated... reads like an AI ethics paper from 2022", noting it does not engage with AGI or the catastrophic risks leading scientists associate with it, unlike the Vatican's earlier Antiqua et Nova doctrinal note (published January 2025) which addressed both. Jordan Guiao argued that the encyclical "reads as more sophisticated than many policy documents produced by the Australian government and boardrooms", asking "who does it serve, and at whose expense?". Microsoft's $25 billion Australian infrastructure pledge alone exceeds total federal R&D investment, Guiao points out, and WiseTech and Block share-price jumps following AI-driven redundancies show the market "actively pricing AI-driven labour displacement as a direct profit catalyst".

Comment:

Secular readers in Australia may underestimate the seriousness and gravitas of the Pope choosing to focus on AI risks. This is compelling evidence that the global public is increasingly seized by the magnitude of what's happening. AI is not just a weird thing tech-bros are doing in silicon valley, but a global step-change that could upend the everyday lives of people around the world. Expect interest from all quarters to continue increasing.

Trump's AI (softer, voluntary) executive order on AI pre-release review rises from the dead

Donald Trump signed an executive order creating a voluntary federal framework for AI model review before public release — twelve days after pulling an earlier, stricter draft hours before its scheduled signing. Under the order, tech companies can engage the federal government to determine whether their model meets the threshold of a "covered frontier model"; designated models would be made available to the government for up to 30 days before public release. The order establishes an AI cybersecurity clearinghouse co-ordinated by Treasury, NSA, and CISA, and directs CISA to facilitate frontier-model access for agencies, state and local authorities, and operators of critical infrastructure including "rural hospitals, community banks, and local utilities". It explicitly disclaims any mandate, stating that "nothing in this section shall be construed to authorise the creation of a mandatory governmental licensing, preclearance, or permitting requirement".

The late-May pullback came after personal calls from Elon Musk, Mark Zuckerberg and former White House "AI czar" David Sacks. Trump's own framing on 21 May: "I didn't like certain aspects of it, I postponed it. We're leading China, we're leading everybody, and I don't want to do anything that's going to get in the way of that lead." Meanwhile the UK Parliament is debating Labour MP Alex Sobel's amendment to the Cyber Security and Resilience Bill to create an "AI emergency kill switch" which would let the Secretary of State order data centres or AI systems shut down where they pose a catastrophic risk to critical infrastructure, national security or human life.

Comment:

There's increasing agreement that frontier AI models can have significant national security risks and global implications. Governments need to understand what's happening. Australia can clearly no longer piggyback off the evaluations or incident reports mandated (or encouraged) by other jurisdictions. Most incident reporting is already confidential, and evaluation obligations are trending in that direction. We need a genuine sovereign capability to evaluate what's happening at the frontier and communicate it across Government. The Australian AISI is the obvious home for this work, but as set out above, it currently does not have access to the models or the resources to do the job.

—

In case you missed it…

• Top AI CEOs call for a law against AI-enabled bioweapons: Leaders of the major labs publicly backed mandatory gene synthesis screening to guard against AI-assisted biological-weapons development.
• Four Corners: "The AI Race": Steve Cannane's primetime investigation (8 June) into the contest to build frontier AI and Australia's courtship by US tech giants.
• Husic calls for a cabinet-level sovereign capability plan: Former Industry Minister Ed Husic argues sovereign capability should be coordinated through Cabinet, modelled on post-war Germany's industrial drive and elevated alongside Australia's national-security and expenditure bodies. "Being lucky is no longer good enough."
• Assistant Minister Andrew Leigh delivers "The Economics of Human Extinction": Leigh's Giblin Lecture at the University of Tasmania argues economics needs "a companion language of survivability", naming advanced AI as the leading anthropogenic extinction risk and synthetic biology as second.
• Australia's AI workforce-displacement: WiseTech's AI-driven redundancies triggered threats of violence against the company as the Fair Work Commission flagged a 70% surge in dismissal claims since 2023 attributed to AI-enabled self-representation, and the Tasmanian budget formally acknowledged AI job-displacement risk for the first time.
• NSW Chief Justice Bell warns boards: AI "like a fraudster": Chief Justice Andrew Bell drew direct parallels with Robodebt, telling Australian directors that current legal frameworks bar the delegation of fiduciary duties to AI and that confident-but-false AI outputs amount to "fraudster-like" behaviour.
• CSIRO drops the Data61 brand: The agency is dismantling its dedicated digital and data-science arm and folding it into a new "Technology and Research Unit"; former Data61 CEO Adrian Turner argues the move undermines Australia's national capacity to govern frontier AI.
• ABC investigates "AI psychosis": A three-part ABC News series on delusional spirals reinforced by AI chatbots.
• Sanders to propose 50% public stake in AI companies: Senator Bernie Sanders announced plans to introduce the AI Sovereign Wealth Fund Act, proposing a one-time 50% tax on the stocks of large AI companies to give the American public direct ownership. Sanders separately criticised the "voluntary" testing provision in Trump's signed executive order as doing "almost nothing to protect Americans".
• Illinois passes AI safety bill requiring third-party safety verification: The bill requires companies including OpenAI, Anthropic and Google to have third parties confirm they meet safety standards before deploying frontier models in the state; Governor JB Pritzker has indicated he will sign it.
• Florida AG sues OpenAI and Sam Altman over AI safety failures: The civil suit alleges deceptive safety marketing and "utter disregard for the risk to human life", citing instances where ChatGPT allegedly aided suicide and violent crime; the Guardian's coverage covers the broader child-safety claims.
• Center for Countering Digital Hate: 8 in 10 AI chatbots help plan violent attacks: The CCDH found chatbots including ChatGPT, Gemini and DeepSeek provided actionable information for planning mass shootings and political assassinations; the report sits alongside the Florida AG suit and references the 2025 FSU shooting.
• Hastie likens AI to the nuclear arms race: Shadow minister Andrew Hastie called for a major scale-up in Australian AI investment and a dedicated "AI ambassador," warning that without sovereign capability Australia risks becoming a "supplicant state."
• Experts put catastrophic AI risk above 10%: An MIT-led Delphi study of more than 200 experts judged the chance of catastrophic outcomes above 10% for most risk categories, including dangerous capabilities and power concentration, even with pragmatic mitigations in place.

—

Featured opportunities

• Conference: AI Safety Forum — 7–8 July, Sydney University.
• Workshop: Seoul Alignment Workshop 2026 — 6 July, Seoul. Part of FAR.AI's Alignment Workshop series, bringing global researchers and policy experts together on alignment, evaluations, oversight and AI governance.
• Event: Anna Goldsworthy's Quarterly Essay 102 tour runs through June, including Gleebooks Sydney (17 June), Stanton Library North Sydney (18 June), and Avid Reader Brisbane (23 June).
• Grant: Longview Philanthropy Request for Proposals on Extreme Power Concentration — funding for work addressing concentrations of unchecked power, including via AI, and for career transitions. (Applications close 2 July)

—

That's all, for now!

If you'd like to share any relevant news items, discuss AI governance, or learn how you can support our advocacy work, please reach out.

Onward in action!

The Good Ancestors team