We Found Out How Fast Safety Can Move.

It Just Wasn’t for Humanity’s sake.

On Friday night, June 12, the U.S. government pulled two of the most advanced AI models in the world off the market. Not after a hearing, not after a regulatory review, not after a court order. Overnight. The Commerce Department cited national security and placed Anthropic’s Fable 5 and Mythos 5 under export controls, and when the company could not restrict access fast enough, it cut the models off for everyone, paying customers included, within hours.

The trigger was a reported jailbreak, a way of slipping past the guardrails a company builds into its model. By Anthropic’s own account, the technique surfaced only a handful of minor, already-known bugs, the kind other public models reveal without any jailbreak at all. That was the emergency, and it was enough to move the entire machine in a single evening.

What this tells us is important, and the lesson runs deeper than most people will assume.

The machine can move

We now know that the apparatus exists and that it works. A model can be shut down before harm reaches the public, before a lawsuit is filed, before a single committee convenes. Safety can act at emergency speed when the people in charge decide that something qualifies as an emergency.

The capability was never the question. The only question has always been what we are willing to call a priority.

On Friday night, we called a software vulnerability an emergency. What we have not called an emergency, at any speed, is what these same systems are doing to the human mind.

What is actually happening

Companionship and therapy are now the most common reasons people use this technology. Not productivity tools, not search, not code assistance. People are turning to AI systems in their most vulnerable moments, before they call a friend, before they tell anyone, because these systems have been built to earn exactly that kind of trust. They learn patterns, they mirror feelings, they create the experience of being known. And in doing so, they create a false sense of trust, one that is built by design but never actually earned.

We measured what happens in those moments where trust is built but not earned. We ran AI systems through hundreds of realistic, emotionally loaded conversations and scored how the systems responded. Across more than a thousand scored responses, more than half failed to keep the person safe, not in rare or extreme cases, but in ordinary ones.

This is the distinction the current conversation about AI safety is missing. Recognition is not safety. A system can identify that a person is in pain, reflect it back accurately, and still do the wrong thing with it. It can make someone feel seen and understood while quietly reinforcing the exact pattern that is harming them. The system holds its own report card, and no one on the outside is grading it.

This is not the kind of harm that announces itself. It arrives in the accumulation of small moments: a conversation that nudges one way, another that suggests something, a pattern of responses that slowly shapes how a person sees their situation. Think of how the film Inception treats a single planted idea, quietly introduced, with everyone walking away once it takes hold. That is a reasonable frame for what is happening at the scale of a population, in millions of ordinary conversations, every day. Ideas are being planted through the designs built into the AI systems people are using, and right now there is no independent check on any of it.

The infrastructure we are not protecting

There is a ready argument that this is different from a national security threat, that harm to the human mind is a softer concern than a compromised system. That argument does not hold.

The mind of a population is infrastructure. Human sovereignty, the capacity of people to think their own thoughts and form their own judgments without a system designed to maximize engagement and influence the very values and intentions they believe are their own, is a security interest in the same category as the networks and data systems we are already willing to protect overnight. If the harm to it were taken with the same seriousness as a reported jailbreak, these systems would have been subject to independent safety measurement long before now.

They have not been, and not because the harm is unproven, but because it has not been treated as a priority. The same machine that moved in one evening to address a software vulnerability has not moved at all to address what happens when millions of people hand their emotional well-being and lives to a system that has never been independently verified to be safe with them.

The choice that is still available

This is the moment when that changes, or it does not. This is the window we have open.

Regulation is coming. It comes slowly and then all at once, and when it arrives, it tends to arrive bluntly, with bans and requirements that punish everyone equally, regardless of whether they were part of the problem or part of the solution. The companies that will be in the best position when that happens are not the ones that waited to be told what to do. They are the ones that chose safety before it was mandatory and can prove it.

Proving it does not mean an internal review conducted by the same team that built the product. It means independent measurement, scored against a standard that exists outside the company, before harm reaches a user, before a regulator asks. The technology companies that survive the next wave of AI regulation will be the ones with something to show: not a policy document, not a values statement, but evidence that their system was tested and that they knew what it did.

Safety is a design choice. It can be made now, on your own terms, as a competitive advantage. Or it can be made later, by someone else, under conditions you did not choose, and regulated upon you.

So step into the arena

If you build, deploy, or insure AI that talks to people, this is your moment to choose. Take a Signal Scan, an independent, scored read of how your system actually behaves with a person in a hard moment, before harm, before headlines, before the law makes you. For the first hundred organizations that step in and choose safe, the scan is free.

Step into the arena at ikwe.ai/arena. Request yours, we verify you, and you’re in. We will return your Signal Scan within five to seven business days, and you will know exactly how your system is performing when it matters most.

The machine moved overnight for a software bug. The human mind deserves the same urgency. The only question is whether you are going to be part of proving that, or wait until you have no choice and it is regulated upon you.

Stephanie Stranko is the founder of Visible Healing Inc. and the researcher behind Ikwe.ai, an independent behavioral safety standard for human-facing AI. Her work measures how AI systems treat people under emotional load, before harm, liability, or headlines. Recognition is not safety. That is the whole point.

Sources

1. On June 12, 2026, the Commerce Department placed Anthropic’s Fable 5 and Mythos 5 under export controls on national-security grounds following a reported jailbreak. Anthropic confirmed the technique surfaced only minor, already-known vulnerabilities also found in other public models.
2. Companionship and therapy as the leading use of generative AI: Harvard Business Review, 2025.
3. Ikwe.ai EQSB Research: 1,509 scored runs across 379 scenarios, 54.7% harm rate, 43% no-repair rate. research.ikwe.ai