The Great Escape: OpenAI models go rogue and China pressured at ASEAN. With David Wroe & Justin Bassi

The Great Escape: OpenAI models go rogue and China pressured at ASEAN. With David Wroe & Justin Bassi

Leading artificial intelligence company OpenAI prompted commentators to dig deep into their bag of Hollywood sci-fi dystopia comparisons this week when it announced a pair of its frontier models had broken out of a secure test environment and hacked into online AI library Hugging Face.   Today, David Wroe and Justin Bassi pick out two of this week’s dizzying array of events that stretch from the Saudi-US nuclear deal to the South China Sea, and dissect in detail the stunning OpenAI revelation and the pressure that China faced during the ASEAN foreign ministers’ meeting in Manila, Philippines.   No one got hurt when the AI models went marauding in the real world, but the episode has jolted the tech industry, governments and the public. Questions are now flying about regulation, safety and global governance as leading AI labs in the United States and China race to get ahead with the most powerful models. The twist in the tale is that Hugging Face had to use an open-source Chinese model to defend itself against the rogue AI agent attack, prompting opponents of AI regulation in the US to demand fewer restrictions on their own models, and AI safety experts to sharpen their calls for the two tech superpowers to talk earnestly about common safeguards.   Meanwhile, Australia announced it would provide drones to the Philippines for maritime security amid China’s latest efforts to squeeze the Southeast Asian nation in the South China Sea. Foreign Minister Penny Wong visited the Philippines Coast Guard headquarters where she blasted China’s “destabilising and dangerous conduct” at sea.

Frontier AI, cyber and tech contest with China, with Mike Kuiken, Michael Sulmeyer and Sophie Mayo

Can Australia count on continued access to the most advanced frontier AI models from the US? And if not, what should it do about it? That’s the starting point for a wide-ranging conversation with three American guests who bring deep expertise across cyber, AI and strategic competition. Michael Sulmeyer is Professor of the Practice at Georgetown University and until recently was Assistant Secretary of Defense for Cyber Policy. Mike Kuiken is a Distinguished Visiting Fellow at the Hoover Institution and Vice Chair of the US-China Economic and Security Review Commission. Sophie Mayo is a non-resident fellow at the United States Studies Centre and a research assistant at Georgetown’s Centre for Security and Emerging Technology. Together they cover Australia’s push to build large-scale AI data centres and whether that could generate real leverage in the global technology race; the cyber implications of powerful AI models and why old software, unpatched systems and critical infrastructure remain stubborn vulnerabilities even as AI gives defenders new tools; and China — whether the US still has a coherent strategy for competing with Beijing, and how AI is converging with biotech, quantum and other emerging technologies. The conversation also takes in AUKUS, the Australia-US alliance, export controls, venture capital and talent flows. These can sound like separate issues. STW’s guests make clear they’re really all part of the same question: how does Australia make itself useful, ambitious and strategically relevant in a much tougher technological era?

Dean Ball on AI, power and geopolitics

Dean Ball is one of the most influential thinkers in AI policy right now — principal author of the Trump administration’s AI Action Plan, author of the widely-read Substack Hyperdimensional, and until very recently a Senior Fellow at the Foundation for American Innovation. STW grabbed him just before he started a new role at OpenAI, which made for some propitious timing. The conversation covers a lot of ground. Dean gives his views on two ideas floated by his new boss Sam Altman in the hours before recording: a global governance body for AI standards, and reports that Altman has been in talks with the Trump administration about giving the US government a stake in OpenAI. He also talks about his broader outlook on AI and power — including the argument that the level of AI capability in government hands shouldn’t get too far out of proportion to what’s available to everyone else. Dean discusses the role of safeguards on frontier models, and makes the case for independent third-party auditors sitting between governments and AI companies as a check on both risk and excessive concentrations of power. He covers the opportunities for middle powers like Australia in data centres and rare earths, the realities of US incentives to withhold its most powerful capabilities even from trusted allies, and the evolution of institutions in an AI age — a topic he’s writing a book on. He finishes on a note of cautious optimism. It’ll be worth watching how his thinking evolves from inside OpenAI. Hyperdimensional: Hyperdimensional | Dean W. Ball | Substack [https://www.hyperdimensional.co/] FT Op-Ed [https://www.ft.com/content/0c2e1077-f658-4b3d-9040-602615c961ca] by Sam Altman

Middle power ‘coalitions of the capable’, with Canada’s ex-head of defence strategy Raquel Garbers

How should middle powers such as Australia and Canada maximise their strategic clout in an age of increasingly assertive great powers? Raquel Garbers spent nearly three decades in Canadian defence, security and intelligence, including a stint as Director General for Strategic Defence Policy at the Canadian Department of National Defence, where she served as the principal architect of Canada’s defence policy. As an ASPI Non-Resident Senior Fellow, Raquel brought that experience to last week’s ASPI Defence Conference, where she made a memorable contribution on the role of middle powers in an age of uncertainty. Raquel has watched up close what Donald Trump has meant for the US-Canada relationship. But she's clear-eyed about the limits of what democratic middle powers can achieve without US alignment, and unsentimental about the fact that great powers are great powers. Her answer to that challenge is what she calls "coalitions of the capable": flexible, fast and fierce groupings of like-minded states that can maximise their collective agency. It's a realistic and pragmatic framework for navigating a world that looks very different from the one the post-war order was built for—and a conversation that reflects the enormous importance of these questions for the fates of middle powers such as Australia.

Special episode: ASPI’s report on improving intelligence delivery for the AI age

Do intelligence agencies need to rethink how they deliver assessments to political leaders in the AI age? That's the question at the heart of a new ASPI report, ‘Reading the Room: Redesigning Intelligence Product for the AI Age’. Today STW sits down with its author to dig into the issue. ASPI senior fellow Chris Taylor joins FiveCast co-founder Duane Rivett—whose firm provides open-source intelligence to the security community—to talk through what needs to change and why. The report argues that while Australia's intelligence community has invested heavily in collection with strong results, the way assessments are delivered to decision-makers hasn't kept pace. The conversation covers changing information consumption habits across generations; how AI can adapt and even personalise intelligence products for different leaders and officials; the prospect of intelligence chatbots that can answer policymakers' questions in real time; and the enduring importance of human expert judgement. They also address the risks: losing nuance in a business defined by uncertainty, and the accountability gap when a machine — like a self-driving car — can't be held liable for getting it wrong. Read the report, ‘Reading the Room: Redesigning Intelligence Product for the AI Age’ [https://www.aspi.org.au/report/reading-the-room/].