12 June – 19 June 2026
HIGHLIGHT OF THE WEEK
Who gets to pull the plug on frontier AI?
Access to powerful AI systems is no longer just a product feature, a subscription tier, or a matter of corporate safety policy. It is becoming a governance decision. When an AI model is capable enough to support software engineering, scientific research, cybersecurity work, critical infrastructure protection, and potentially harmful misuse, the question is no longer only how well the model performs. The question becomes who should be allowed to use it, under what conditions, and who has the authority to switch it off.
That question moved from theory to practice over the past week, after Anthropic disabled access to Claude Fable 5 and Claude Mythos 5 in response to a US government export-control directive. The timing made the case especially striking. Only days earlier, Anthropic had launched Fable 5 as its most capable generally available model to date, alongside Mythos 5, a more powerful version reserved for selected cyber defence and critical infrastructure partners.
At first, the launch looked like a model governance experiment. Fable 5 was made broadly available, but with additional safety classifiers designed to detect potential misuse, including attempts to bypass safeguards. Anthropic also introduced a fallback system that routed some high-risk sessions to a less capable model. Mythos 5, meanwhile, used the same underlying model, but with some safeguards lifted for trusted users working in cyber defence and critical software infrastructure.
In other words, Anthropic was trying to answer one of the hardest questions in frontier AI deployment: how can companies give users access to more capable systems while reducing the risk that the same systems are misused in sensitive domains? Its answer was tiered access, safety filters, model routing, testing, and trusted partnerships.
Then came the cutoff.
What followed was more than a technical compliance problem. According to Anthropic, the US directive required the company to suspend access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. The company said the practical result was that it had to remove access to both models for all customers to ensure compliance, while other Anthropic models remained available.
Why is this happening? The immediate reason centres on concerns that Fable 5 could be jailbroken. Anthropic said it understood that the government had become aware of a method for bypassing the model’s safeguards, but also said the directive did not provide specific details of the national security concern. The company argued that the technique it reviewed appeared to identify a small number of previously known minor vulnerabilities, which could also be found using other publicly available models.
That disagreement matters because it points to a much larger governance problem. Governments clearly have a legitimate interest in preventing the misuse of powerful AI systems, especially when they may affect cybersecurity, biological research, defence, or critical infrastructure. But companies, users, researchers, allies, and civil society also have an interest in knowing when a deployed model can be recalled, what evidence is required, and what transparency or due-process safeguards apply.
If the standard for suspending access is too weak, frontier AI deployment becomes unpredictable. If the standard is too high, genuinely dangerous systems may remain available for too long. The difficult part is building a system that can respond quickly in real emergencies without allowing access to advanced AI to become an unclear political decision.
The episode also gives new force to the debate on AI sovereignty. Governments outside the United States have spent the past year asking whether they can rely on AI systems controlled by a small number of foreign companies. The Fable 5 and Mythos 5 restrictions made that question less abstract. If foreign users, trusted partners, and even allied institutions can lose access overnight due to a US national security decision, then reliance on frontier AI providers becomes a strategic vulnerability.
That does not mean every country can build its own frontier models. AI sovereignty is easier to demand than to deliver. Training and deploying advanced systems requires chips, data centres, energy, skilled researchers, secure cloud infrastructure, capital, and access to global supply chains. For many governments, the real choice may not be between full dependence and full sovereignty, but between managed dependence, trusted access, regional capacity, and strategic resilience.
For the United States, there is also a trade-off. Broad restrictions may reduce certain security risks, but they can also erode trust with allies, disrupt research cooperation, and encourage other countries to accelerate the development of domestic alternatives. AI leadership is not only about having the most capable models. It also depends on whether partners believe that access to those systems is predictable, lawful, and governed by clear rules.
The Fable 5 saga shows that frontier AI is becoming more like strategic infrastructure than ordinary software. Access decisions can affect cybersecurity teams, research institutions, public agencies, private companies, and international partnerships. Safety cannot rely solely on company policy, but government intervention also cannot rest on vague national security claims.
The bigger question is whether societies can design rules for frontier AI access that are fast enough to address real risks, transparent enough to build public trust, and stable enough to support international cooperation. If powerful models can be launched one week and pulled back the next, the politics of AI will increasingly revolve not only around what these systems can do, but around who gets to decide who can use them.
IN OTHER NEWS LAST WEEK
AI governance and digital security
G7. Leaders of the G7 nations gathered in Évian-les-Bains, France, for the 2026 summit, with critical mineral supply chains, online child safety and international solidarity among the key issues on the agenda. France placed balanced economic growth and international cooperation at the centre of this year’s discussions, while leaders also addressed geopolitical tensions, support for Ukraine, regional security and the Middle East. The digital policy dimension was visible through debates on resources needed for semiconductors and digital infrastructure, as well as the growing connection between economic security, technology governance and geopolitical stability.
The EU. The European Commission welcomed the new G7 Cybersecurity Working Group Declaration, adopted under France’s G7 Presidency. The declaration calls for stronger international cooperation on cybersecurity challenges linked to quantum computing, AI, telecommunications infrastructure and the protection of small and medium-sized enterprises. It treats the transition to post-quantum cryptography as an urgent priority, while also recognising that AI can both strengthen and threaten cybersecurity through AI-enabled attacks, model manipulation, data breaches and software vulnerabilities.
G7. Digital and technology ministers adopted the group’s first shared set of principles for protecting children and young people online. The agreement addresses risks linked to harmful content, exploitation and AI chatbots, while calling for stronger digital literacy, safety measures built into digital services from the start, and effective age assurance. Ministers also called for better access to data and research on how digital services affect children’s well-being, alongside closer cooperation among platforms, researchers, families and governments.
The EU. The European Commission also welcomed the G7’s adoption of online child protection principles, noting that they reflect approaches already present in EU initiatives such as the Digital Services Act, the Better Internet for Kids Strategy and the AI Act. The framework promotes safety-by-design measures, privacy-conscious age assurance tools, stronger protections against harmful and illegal content, parental controls and digital literacy. It also links child protection with privacy, fundamental rights and access to digital opportunities, showing how online safety policy is moving beyond content removal alone.
UNIDIR launches AI peace and security platform in Geneva
The UN Institute for Disarmament Research, together with Switzerland and Pakistan, hosted a pre-launch briefing for its Centre of Excellence on AI, Peace and Security at the Palais des Nations in Geneva. UNIDIR said the centre is being established at a critical moment for global AI governance, as AI increasingly reshapes international peace and security dynamics. The centre is intended to serve as a permanent platform for consolidating knowledge, connecting stakeholders and maintaining continuity between multilateral processes and global discussions on AI and international security.
The platform will support more coherent international AI governance while promoting inclusive global engagement. It will provide practical, evidence-based policy guidance, resources and capacity-building support for actors working on AI in peace and security contexts. The initiative also points to Geneva’s growing role as a hub where AI governance, security policy, disarmament debates and multistakeholder dialogue increasingly overlap.
Children’s online safety under UK scrutiny
The UK Government Office for Science highlighted the need for evidence-led policy on children’s online lives, warning that digital technologies bring both benefits and risks while long-term evidence remains limited. The article noted that 97% of UK teenagers aged 13 to 15 now own a mobile phone, while almost one-fifth of children aged three to five also own one. Children use digital platforms to maintain friendships, access communities, and find support, while AI tools are increasingly used for learning, schoolwork and, in some cases, well-being and emotional regulation.
However, the government adviser warned that children face risks including harmful content, cyberbullying, privacy breaches, false or misleading information, unlimited scrolling, personalised algorithms and other engagement-maximising design features. The article also cautioned that the use of AI should not prevent children from developing skills such as written expression and critical thinking. The debate shows why child online safety is increasingly being framed not only as a platform moderation problem, but as a question of design, evidence, education and long-term development.
IWF warns of AI-generated child sexual abuse risks
The Internet Watch Foundation welcomed Pope Leo XIV’s reflections on AI, arguing that AI systems must be developed with stronger safeguards to protect children from abuse. The organisation warned that AI is being used to generate highly realistic child sexual abuse images and videos at scale, saying the number of AI-generated child sexual abuse videos it identified in 2025 increased by more than 260%, with nearly two-thirds falling into the most severe category of abuse.
The IWF also raised concerns about AI-nudification tools, which can generate realistic sexualised images of children and other individuals. Following the Child Dignity in the Artificial Intelligence Era conference in Rome, the organisation joined more than 100 organisations and individuals in supporting calls for a global ban on such tools. It also called for stronger regulation, global alignment and enforceable safety-by-design standards to prevent the creation and spread of AI-generated child sexual abuse images and videos.
The EU extends cyber crisis support to Ukraine
Ukraine can now activate emergency EU cyber support during significant or large-scale cybersecurity incidents after the Council of the EU approved its inclusion in the EU Cybersecurity Reserve. The reserve, managed by the European Union Agency for Cybersecurity, provides incident response services from trusted private-sector providers to help contain and mitigate major cyber incidents.
The European Commission said the decision reflects closer EU-Ukraine cooperation and forms part of wider efforts to strengthen preparedness, rapid response and shared expertise against evolving cyber threats. The move also aligns with the EU’s strategic digital partnership agenda and follows Moldova’s inclusion in the Cybersecurity Reserve in 2024 under the Cyber Solidarity Act. Ukraine’s inclusion shows how the EU is using cyber crisis mechanisms not only for internal resilience but also for strategic partnerships in a conflict-affected digital environment.
Latin America looks to Europe’s AI rulebook
Countries across Latin America are increasingly drawing on the EU’s risk-based approach as they develop AI governance frameworks, according to an International Bar Association analysis. Chile, Peru, Brazil and El Salvador have introduced or advanced AI-related legislation in recent years, with several initiatives focusing on risk classification, fundamental rights and accountability. Peru implemented an AI law in 2023 and updated it in 2025, while Brazil has considered similar draft legislation, and El Salvador enacted an AI law in 2025.
Chile has become one of the region’s most active AI policy actors, helping launch Latam-GPT, an open-source language model trained on Latin American data, while considering an AI bill that would classify systems according to risk. The proposal includes stricter obligations for high-risk uses, regulatory sandboxes and measures intended to support innovation and smaller businesses. However, legal experts warned that EU-style requirements on risk governance, technical documentation, testing and incident reporting could strain regulators and smaller companies with limited resources.
Swiss parliament weighs AI apps in media copyright bill
Swiss lawmakers want the government to examine whether AI applications should be covered by a media copyright bill that requires online services to compensate publishers for displaying extracts of newspaper articles. The Swiss Senate unanimously referred the bill back to the federal government, while the House of Representatives had already approved the request in March by 157 votes to 29, with two abstentions.
Why does it matter? The debate shows how AI-powered search tools, chatbots and assistants are being pulled into older disputes over platform power, copyright and publisher compensation. As AI services become gateways to information, governments are asking whether existing media copyright rules remain fit for a digital environment shaped not only by search engines and social platforms but also by AI applications.
MIT researchers target AI data centre cooling
A startup founded by MIT researchers has developed a nuclear-inspired cooling system to improve data centre energy efficiency while reducing water consumption. The technology targets one of the fastest-growing sources of electricity demand, as the rapid expansion of AI infrastructure drives increased computing requirements.
Ferveret’s system uses a specialised liquid to immerse servers, replacing traditional air-based cooling methods that account for a significant share of data centre energy consumption. Its Adaptive Phase Cooling approach improves heat transfer through controlled bubble formation, increasing efficiency while reducing reliance on water-intensive cooling systems. The company reports computational efficiency gains of up to 15% compared with existing liquid cooling technologies and says its modular design could help data centres operate more effectively in regions with limited water resources.
LAST WEEK IN GENEVA
Geneva placed AI, security and ethics at the centre of international security diplomacy this week. From 15 to 17 June, UNODA convened informal exchanges on AI in the military domain and its implications for international peace and security, following the UN General Assembly’s request for states and stakeholders to discuss observations from the Secretary-General’s report, opportunities and challenges, emerging normative proposals and possible next steps.
The exchanges brought together states, observer states, international and regional organisations, the ICRC, academia, civil society, the scientific and technical community and industry. Discussions and side events addressed the responsible development and use of AI in military decision-making, trustworthy AI in the military domain, human rights across the life cycle of military AI, assurance and accountability, the role of private actors, weaponisation of data-intensive technologies, and the nuclear dimension of AI.
The Geneva week then continued with UNIDIR’s Global Conference on AI, Security and Ethics 2026, held on 18 and 19 June in Geneva and online. The conference brought diplomats, policymakers, military representatives, researchers, industry, civil society and international organisations together to examine how AI is reshaping national, regional and global security environments. The programme focused on themes ranging from algorithmic bias, dual-use technologies, agentic AI, testing and assurance, and AI in cyber defence to counter-AI capabilities, infrastructure dependencies, public-private responsibility chains and human-centred approaches to defence AI.
Together, the UNODA exchanges and UNIDIR conference showed how Geneva’s AI diplomacy is moving beyond broad concern about emerging technologies. The agenda is increasingly about how to translate international law, responsible AI principles, military practice, technical testing and multistakeholder expertise into workable governance for high-stakes systems. As AI becomes embedded in defence, cybersecurity and strategic infrastructure, Geneva is becoming one of the key spaces where the security implications of AI are debated not only as future risks, but as present governance challenges.
LOOKING AHEAD
The Developments in Africa cyber diplomacy: Continental, regional, and national initiatives webinar will take place online on 24 June, organised by Diplo. The event will examine the evolving landscape of cyber diplomacy across Africa, focusing on continental, regional and national initiatives. Discussions will cover key players, stakeholders and frameworks shaping cyber diplomacy on the continent, including African Union initiatives and the African Digital Compact, as well as national strategies for addressing cyber threats and strengthening digital cooperation. The webinar aims to bring together policymakers, cybersecurity experts, researchers, civil society and private sector stakeholders to exchange lessons, identify best practices and support secure and inclusive digital transformation across the continent.
UN Open Source Week 2026 will take place from 22 to 26 June at UN Headquarters in New York, co-hosted by the UN Office for Digital and Emerging Technologies and the UN Office of Information and Communications Technology. The event is framed as a global convening on open-source technologies and digital cooperation, with a programme covering a Community Hackathon, open source and AI, Digital Public Infrastructure Day, Open Source Programme Offices for public-interest objectives and community-led events. The week will explore how open source can support the SDGs, the Global Digital Compact, digital public infrastructure, responsible AI development and practical collaboration among UN entities, member states, technical communities, civil society and the private sector.
READING CORNER
Agentic AI systems that can plan, decide and act autonomously across platforms simultaneously trigger obligations under data protection, competition, financial services and online safety law, exposing gaps that sector-based regulation…
Anthropic’s Claude Fable 5 went public, then vanished 72 hours later. Inside the sudden U.S. government shutdown and what it means for the future of AI power.
DON’T MISS
Follow the road to the 2027 Geneva AI Summit, a global journey shaping AI’s future. Bookmark the page and register for updates to stay informed and involved every step of the way to Geneva 2027.
