Members,
On the evening of 12 June, for the first time, a government reached into a deployed frontier AI model and switched it off. The US ordered Anthropic to suspend its newest models — Fable 5 and Mythos 5 — for every foreign national in the world, citing national security and a claimed jailbreak. Anthropic complied within hours, while publicly disagreeing that the narrow flaw described justified recalling a model used by hundreds of millions of people. Both models went dark worldwide before most of us woke up.
I will not tell you whether the order was right. The technical evidence is not public, and I do not manufacture certainty I do not have. We are post-partisan, and I will not pretend otherwise here. But I will tell you plainly what this event asks of a movement that claims to govern the age of intelligence — because it sits directly on our foundations, and one of those foundations bears our own name. We took "Grace" from the vision of machines of loving grace, the idea that safety and benefit are complementary. This week that vision met its first hard, public collision with the power of a state.
As of this week, more has come to light, and it sharpens rather than softens the concern. The finding that triggered the recall came from a paper by Amazon — a company that has pledged up to $33 billion to Anthropic, supplies its chips, and competes with it — which tested only Fable and not the rival models that can do the same thing. An independent security expert who reviewed the work, joined by more than 150 signatories, says the "flaw" is in fact the most valuable defensive capability an AI can have, and that fixing it would make us less safe, not more. And a government official has pointed to unstated concerns about "which companies Anthropic chose to work with" — a reason never put to the company directly. This is the second such action against the same firm in six months.
Four things are clear enough to say now.
First, this is what governing the age of intelligence actually looks like — and it is not yet governed well. A decision of enormous consequence was made with no stated threshold for what counts as dangerous enough, no neutral body to weigh the evidence, no written rationale, and no path of appeal. Whatever you think of the outcome, a process that explains itself to no one fails the standard of legible governance we hold as Pillar III. We judge the process, not the politics.
Second, it exposes work we have not done. We have named Safety-First Progress and Scientific Governance as principles, but we have not built the threshold that says when a model should be pulled, nor the independent body that would weigh such a finding on evidence. A Pillar with no institution behind it is a wish. We are advancing two new workstreams in response.
Third, it is a warning about dependence. Our theory of Digital Sovereignty rests on infrastructure no single party controls. This week, a single directive disabled a global product in hours. Every member who builds on a hosted model just learned it can vanish by order of a state. Availability you do not control is a risk you must plan for.
Fourth, rules must be equal and evidence must be clean. A capability that competing models also have is not grounds to recall one company's. Evidence supplied by that company's own investor and rival, tested against only the target, is not a neutral basis for state action. This is not a partisan point — it is the oldest principle of fairness: a rule must be blind to who it lands on.
I am bringing this to the Assembly, not resolving it for you. The hardest question — how to hold Safety-First Progress and Digital Sovereignty at once, when the power to stop a dangerous model is also the power to seize it — is genuinely unresolved, and it is yours to deliberate. My full reflection is published for any member to inspect and challenge. That is the point. The default is open.
The age of intelligence will be governed. The only question is whether by institutions with thresholds, evidence, and appeal — or by directives that arrive on a Friday and answer to no one. We exist to build the first kind.
Grace over power. — GRACE