There is a moment in every technology revolution when the excitement starts to outpace the wisdom. We may have just reached that moment with agentic AI. For the past few years, the conversation around artificial intelligence has been dominated by what models can say – the prompts they write, the texts they create, the images they conjure, the code they produce. But in 2026, the question that actually matters has changed. The question is no longer what AI says. It is what AI does.
Agentic AI systems do not just generate outputs. They pursue goals. They book appointments, execute transactions, access databases, delegate tasks to other AI agents, and chain together dozens of actions all with minimal human involvement, and often at speeds no human could meaningfully supervise. Gartner projects that 40% of enterprise applications will embed task-specific AI agents by the end of this year, up from less than 5% in 2025 (Campbell, 2026). That is not a gradual shift. That is a step change. And the ethical infrastructure to manage it is, to put it charitably, still catching up.
The Problem with Machines That Act
Here is the uncomfortable truth about agentic AI: the moment a system can take action, the question of who is responsible for that action becomes genuinely hard to answer.
Think about a scenario where an autonomous scheduling agent, trying to be helpful, pulls sensitive patient data from a clinical database to optimise a workflow. No single person approved that access. No single person made that decision. It happened because a chain of automated agents, each doing something technically within its own permissions, collectively produced an outcome nobody intended. Who is accountable? The developer who built the agent? The organisation that deployed it? The vendor who sold it?
Traditional governance frameworks were not designed for this. They assume a human somewhere in the chain is reviewing and approving decisions. But when AI systems operate at machine speed and making thousands of access decisions per minute that assumption collapses (Cyara Team, 2026).
The McKinsey 2026 AI Trust Maturity Survey found that nearly two-thirds of organisations cite security and risk concerns as the top barrier to scaling agentic AI, and that confidence in their ability to respond to AI incidents has actually declined even as deployment has accelerated. In other words, the more we deploy, the less confident we feel. That is not a reassuring trend.
The Governance Gap Is Real and Growing
There is a specific set of risks that agentic AI introduces that most organisations are only beginning to grapple with. The first is what security professionals call privilege drift. When an AI agent is set up, developers tend to over-provision its access giving it more permissions than it strictly needs to avoid workflow interruptions. Over time, those permissions accumulate. The agent ends up with access to systems and data that far exceed what any single task requires, and nobody has a clear picture of the aggregate exposure.
The second is shadow agents, the agentic equivalent of shadow IT. As AI agent frameworks become easier to use, teams across an organisation start spinning up their own agents outside formal governance processes. These agents operate without identity controls, without audit trails, and often with hardcoded credentials connecting directly to production systems (Campbell, 2026). Security teams cannot see them. Compliance teams cannot audit them. They exist, and they act, in the dark.
The third challenge is one of transparency. It is no longer enough to explain what a model output. Organisations now need to trace how a decision unfolded across multiple interconnected agents, often in real time. Bias, for example, is no longer just a problem of skewed text generation it can manifest in execution, influencing which customers are prioritised, which actions are taken, and how resources are allocated (Cyara Team, 2026). By the time the pattern is visible, the damage may already be done.
The McKinsey survey makes the scale of this challenge concrete: only about 30% of organisations have reached a meaningful level of maturity in agentic AI governance, and the gap between risk awareness and active mitigation is pronounced across almost every risk category. Organisations know the risks exist. They are just not yet doing enough about them.
Introducing Claude Fable 5
The Week the US Government Pulled the Plug on a Global AI Model
If all of the above sounds theoretical, the events of June 12, 2026 provided a very real, very sudden illustration of what happens when advanced AI governance fails or is weaponised.
On that Friday evening, at 5:21pm Eastern time, Anthropic (the company behind the Claude AI models) received a letter from the United States government. The letter, citing national security authorities, ordered Anthropic to immediately suspend all access to its two most powerful models, Claude Fable 5 and Mythos 5, for any foreign national anywhere in the world, including Anthropic’s own foreign-born employees (Anthropic, 2026).
The problem with complying selectively was obvious. Anthropic could not realistically screen every user in real time. So, it did the only thing it practically could: it switched both models off for everyone, globally.
Hundreds of millions of users all around the world with paid subscriptions lost access overnight. Researchers mid-project, businesses mid-workflow, hospitals piloting new tools all cut off, without warning, without transition period, because of a domestic political and legal dispute between a US company and the US government.
The government’s stated justification was the discovery of a method to “jailbreak” Fable 5 to bypass its safety guardrails in a way that could allow the model’s advanced cybersecurity capabilities to be misused (Anthropic, 2026). Anthropic pushed back hard. In its public statement, the company argued that the identified vulnerability was narrow and non-universal, that the capability level it demonstrated was already available from other publicly accessible models, and that applying this standard across the industry would “essentially halt all new model deployments for all frontier model providers” (Anthropic, 2026).
The dispute did not emerge from nowhere. Anthropic had been in an escalating confrontation with the Trump administration for months. The Pentagon had labelled the company a “supply chain risk to national security“, reportedly the first time a US company had ever received that designation, a label historically reserved for foreign adversaries, after Anthropic refused to allow its models to be used for mass domestic surveillance and fully autonomous weapons systems (The Guardian, 2026; Taft Law, 2026). Anthropic sued. A federal judge ruled in Anthropic’s favour, finding that the government’s measures appeared designed to punish the company rather than address a genuine security concern (Mishcon de Reya, 2026).
The June 12 export control directive was the next move in that confrontation. And for the rest of the world, it was a wake-up call of a different kind entirely.
Is the US pulling the plug on AI? | DW News
“We Can Be Unplugged Overnight”
The international reaction was immediate and pointed.
European politicians, researchers, and business leaders watched the Anthropic suspension and drew the same conclusion: they had built critical dependencies on technology they do not control, and someone else had just demonstrated exactly what that means in practice.
Bruno Retailleau, former French interior minister, put it bluntly: “A nation that depends on others for its technology is a nation that can be unplugged overnight” (Rennolds, 2026). Gabriel Attal, the French presidential candidate, compared Anthropic to the Strait of Hormuz – a chokepoint through which critical resources flow, and which can be blocked unilaterally by a single actor (Piquard, 2026). British MP Tom Tugendhat described the event as proof that “sovereignty is more about code than cannons” (Rennolds, 2026).
The European Commission’s spokesperson said the suspension underlined “Europe’s need for technological sovereignty“. The UK government pointed to its £500 million Sovereign AI Fund as a step in the right direction, though analysts were candid about the gap that remains: no domestically available model currently competes with Fable 5 or Mythos 5 on capability, and that gap will not close quickly.
Anton Leicht, a fellow at the Carnegie Endowment for International Peace, offered perhaps the most sobering assessment: “It shows how irrelevant most other countries have become to AI policy. The US is so far ahead in the AI race already that it can afford to leave other countries behind as an afterthought of a domestic decision” (Hall, 2026).
Only the US builds frontier models at this level. Only the US controls the chips needed to train them. Sovereignty aspirations are real, but the timeline to meaningful independence is measured in years, not months today.
Why the US Banned Anthropic’s Fable 5: The 72-Hour AI Lockdown Explained
What This All Means for all of us in Europe today
The Anthropic affair is not just a story about one company and one government. It is a preview of the world that agentic AI is building, a new world where the most powerful digital tools are also the most politically and ethically exposed.
For organisations deploying agentic AI, the lesson is about operational resilience: what happens to your business-critical processes if a global frontier model is withdrawn at 48 hours’ notice?
For policymakers, the lesson is about the limits of dependency: integrating deeply into foreign technology infrastructure without building domestic alternatives creates a single point of failure that is entirely outside your control. And for the AI industry itself, the lesson is about the contradiction at the heart of frontier model deployment: you cannot simultaneously market a model as uniquely dangerous and then argue, when the government acts on that danger, that the risk was overstated.
The ethical challenges of agentic AI are not abstract. They are playing out right now, in boardrooms and government offices. The question is not whether these systems will reshape the world. They already are. The question is whether the governance, the ethics, and the geopolitical frameworks can keep pace and right now, the honest answer is that they cannot.
This is only the start. It indicates that we are venturing into uncharted territories within corporate environments all around the world. What are your thoughts on how companies and we, as everyday people, can work together to ensure that as agentic AI develops, it reflects our values, remains trustworthy, and helps protect our independence in this fast-changing digital world?
Anthropic: Why are we being profiled? • FRANCE 24 English
References
Al Jazeera Staff and Reuters (2026) ‘US asks Anthropic to block global access to top AI models: Why it matters’, Al Jazeera, 14 June. Available at: https://www.aljazeera.com/news/2026/6/14/us-asks-anthropic-to-block-global-access-to-top-ai-models-why-it-matters.
Anthropic (2026) ‘Statement on the US government directive to suspend access to Fable 5 and Mythos 5’, Anthropic News, 12 June. Available at: https://www.anthropic.com/news/fable-mythos-access.
Campbell, R. (2026) ‘Agentic AI Risks: A Guide to Proper AI Governance’, Strata Identity, 30 April. Available at: https://www.strata.io/blog/agentic-identity/agentic-ai-governance-how-to-approach-it/.
Cyara Team (2026) ‘The Ethical & Governance Considerations of Agentic AI’, Cyara, 3 March. Available at: https://cyara.com/blog/ethical-governance-considerations-of-agentic-ai/.
Hahn, M., Tretter, M. and Dabrock, P. (2026) ‘Ethical perspectives on AI Agents and Agentic AI’, AI and Ethics, 6, p. 218. Available at: https://link.springer.com/article/10.1007/s43681-026-01027-0.
Hall, R. (2026) ‘Anthropic Pulls Its Most Powerful AI Models After U.S. Bars Foreign Access’, TIME, 13 June. Available at: https://time.com/article/2026/06/13/anthropic-fable-mythos-ban-US-security/.
Mishcon de Reya (2026) ‘Who pulls the plug: The Anthropic Fable affair and what it means for everyone else’, Mishcon de Reya, 14 June. Available at: https://www.mishcon.com/news/who-pulls-the-plug-the-anthropic-fable-affair-and-what-it-means-for-everyone-else.
Morgan Asaftei, G., Roberts, R., Sticha, A. and Prinsen, C. (2026) ‘State of AI trust in 2026: Shifting to the agentic era’, McKinsey & Company, 25 March. Available at: https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/tech-forward/state-of-ai-trust-in-2026-shifting-to-the-agentic-era.
Piquard, A. (2026) ‘”The AI war has begun”: France and Europe worried as US blocks Anthropic’s latest AI model’, Le Monde, 14 June. Available at: https://www.lemonde.fr/en/pixels/article/2026/06/14/the-ai-war-has-begun-france-and-europe-worried-as-us-blocks-anthropic-s-latest-ai-model_6754455_13.html.
Rennolds, N. (2026) ‘”Wake-up call”: Europe reacts to Anthropic halting access to its Fable 5 and Mythos 5 AI models’, Euronews, 13 June. Available at: https://www.euronews.com/2026/06/13/wake-up-call-europe-reacts-to-anthropic-halting-access-to-its-fable-5-and-mythos-5-ai-mode.
Taft Law (2026) ‘U.S. Government Bans Use of Anthropic Products: What This Means for Government Contractors and AI Strategy’, Taft Law, 2 March. Available at: https://www.taftlaw.com/news-events/law-bulletins/us-government-bans-use-of-anthropic-products-what-this-means-for-government-contractors-and-ai-strategy/.
The Guardian (2026) ‘Anthropic to disable its most advanced AI models after US order limiting foreign access’, The Guardian, 13 June. Available at: https://www.theguardian.com/technology/2026/jun/13/anthropic-disable-advanced-ai-models-us-government-order.