A single government letter can now switch off a frontier AI capability overnight. That is the central lesson of Anthropic's sudden removal of its Fable 5 and Mythos 5 models, and it is the clearest warning to date about AI supply chain and concentration risk. The tools that businesses increasingly build on can be reclassified, restricted, or withdrawn with little notice. Strategies built around a single provider no longer suffice. Nor does a nominal backup.
What happened
On Friday 12 June 2026, Anthropic received a letter from the US Commerce Secretary ordering it to suspend access to Fable 5 and Mythos 5. The directive cited national security concerns. It applied to any foreign national, inside or outside the United States, including Anthropic's own employees.
The government's stated concern relates to a reported method of bypassing Fable 5's safeguards. Anthropic has disputed this publicly, describing the method as narrow and not unique to its models. The firm disabled both models worldwide that evening to comply. Its other models were unaffected.
The timing is striking. Fable 5 had launched to the public only three days earlier, as the firm's most capable generally available model. Mythos 5 had been restricted to a small, vetted group of cyber security partners under Project Glasswing. Anthropic has disagreed with the directive's grounds and has stated it is working to restore access.
Three conclusions for any organisation deploying AI
For all organisations meaningfully deploying AI, three conclusions follow.
1. Vendor concentration is a systemic risk. A handful of foundational models sit underneath critical workflows across most large organisations. Should those models become unavailable, the consequences could be wide-ranging.
2. Regulation now moves faster than procurement. Enterprise procurement cycles once absorbed data privacy and cloud computing rules with time to spare. They have now fallen behind. Regulatory decisions on AI increasingly arrive with only weeks of notice.
3. Diversification alone does not ensure resilience. Swapping AI providers means complex structural change to applications, processes, and controls. It is not simply a contractual switch.
A precedent, not an incident
The specifics of this directive will matter less than the precedent it sets. The US government has shown that a foundational AI capability can be removed from commercial circulation at speed, with limited regard for the disruption caused to users.
Earlier technology controls were mostly confined to specific jurisdictions. AI restrictions are likely to resemble ITAR in their complexity. Re-export rules may extend to derivative products, fine-tuned models, and hosted inference, rather than only the underlying weights. The private sector will push back and work around the edges. In the meantime, these restrictions sit on top of existing operational concerns as a regulatory and compliance burden.
The systemic story is bigger. AI supply chains are being pulled in incompatible directions by US export controls, the EU AI Act, and China's generative AI rules, among other national regulations now emerging. A global organisation's AI implementation must increasingly satisfy regimes that are starting to contradict one another. Those regimes may change at unprecedented pace.
Data privacy and cloud regulation should be seen as warm-ups, not true precedents. AI resilience and compliance will be harder for three reasons. Behavioural lock-in runs deeper, because prompts, evaluations, and fine-tuning are built around the quirks of a specific model. A capability cliff exists at the frontier, where top-tier models are far less interchangeable than hyperscale cloud is today. And the compute underneath—chips, data centres, and energy—is now geopolitically contested in a way cloud infrastructure was not at the outset.
The instinctive response is to line up an alternative provider, perhaps a European or Chinese model as a regional hedge. The instinct is understandable. It is also likely to prove insufficient.
Migrating workloads between AI providers is closer to changing a hyperscale provider than to switching an application. Applications and business processes come to rely on the specific behaviour of a given model, including the non-deterministic outputs of agentic systems. Traditional disaster-recovery and resilience playbooks assume technology is interchangeable. AI models are highly unlikely to be so as they become more deeply embedded.
Regional diversification is not free of cost either, even where it is operationally acceptable. European and Chinese alternatives will position themselves as hedges. Each comes with a regulatory and data-handling profile that interacts with the obligations a global company already carries. Vendor choice is now as much a question of resilience and jurisdiction as of capability and cost.
Frontier capability where it matters, not everywhere
There is a quieter consequence to the Anthropic episode. The US restrictions target frontier models, where access is now becoming a function of nationality, clearance, and jurisdiction. Applied more broadly, these restrictions risk creating a two-speed system. A narrow set of firms would retain reliable access to top-tier capability. Others would operate on whatever is left, with little notice of when the line moves.
Organisations should respond on two fronts: resilience and strategic use. Not every workflow needs a frontier model. A significant share of production use cases can be served well by smaller, open, or locally hosted models—often at lower cost, with reduced regulatory exposure and greater control. Being deliberate about where frontier capability is required, and where it is not, is among the strongest mitigations available to organisations that may not have consistent access to the frontier.
Four questions to ask in every organisation
Boards and executive committees do not need the technical answer. They need to ensure their organisations can answer four questions credibly and consistently.
1. Where are we concentrated? Which products, processes, and decisions depend on a single model or vendor? What would disruption actually cost in revenue, customer impact, and regulatory exposure?
2. How quickly could we substitute? If a primary model were withdrawn tomorrow, how long would a switch take? Which applications and business functions would break or degrade in the interim?
3. Are we placing AI by outcome? Rather than planning around the heaviest workloads, identify where outputs feed a regulated or high-stakes decision—credit, hiring, medical triage, or executive sign-off. Then judge whether those use cases need a frontier model or could be served by a smaller, controlled alternative.
4. Is resilience built in, or assumed? Do we have the testing, monitoring, fallback patterns, and documentation to make a switch viable under pressure? Or are we relying on the absence of a shock?
Resilience as advantage
AI adoption should not be slowed by fear of instability. There is a competitive cost to standing still, and the firms that move fastest will be those that have planned for change in their AI supply chain. Handled well, resilience becomes a source of advantage. It buys pricing leverage, regulatory optionality, the discipline of matching the model to the job, and a moat that locked-in competitors will not have.
Find out how Control Risks can help your organisation manage AI risk with confidence.