By Jeffrey T. Donner, Esq.

June 18, 2026

The public argument over Anthropic’s Fable 5 and Mythos 5 models is not just another technology story. It is a glimpse into the world we are already living in.

According to Anthropic, the United States government issued an export-control directive requiring the company to suspend access to Fable 5 and Mythos 5 by foreign nationals. Anthropic says the order was so broad that, as a practical matter, it had to disable access for all customers. The stated concern appears to involve a possible jailbreak of Fable 5 — meaning a way to bypass the model’s safety controls — and the model’s ability to identify software vulnerabilities.

That sounds dramatic. It also sounds confusing. Does this mean AI can break into government computers? Does it have the President’s password? Does it have nuclear launch codes? Is this the beginning of the robot takeover?

No. At least not in that Hollywood sense.

But it does mean something serious. It means the United States government is beginning to treat access to the most advanced AI models as a national-security issue. That alone is a major development.

This Is Not Skynet

I never found the “Skynet” version of artificial intelligence risk very convincing. That is the Terminator idea: a military AI becomes self-aware, takes control of weapons, launches nuclear war, builds robots, and tries to exterminate humanity.

That is not what happened here.

Anthropic’s models did not escape from a lab. They did not seize control of servers. They did not develop arms and legs to stop humans from unplugging the machine. The government told Anthropic to shut down access, and Anthropic shut down access. That is not a story of AI escaping human control. It is a story of ordinary institutional control: government, companies, servers, cloud access, compliance departments, export-control law, and lawyers.

So if the question is whether this episode proves the hard singularity has arrived, the answer is no.

The “hard singularity” is the idea that AI becomes broadly smarter than human beings, strategically autonomous, self-improving, and no longer controllable by human institutions. That is not what this incident shows.

But there is another version of the singularity that is much more plausible and much more immediate.

Call it the soft singularity.

The Soft Singularity

The soft singularity is not when AI becomes king. It is when AI becomes the clerk, associate, analyst, researcher, coder, auditor, strategist, and consultant sitting next to every decision-maker.

That has already happened.

Lawyers use AI to draft emails, summarize documents, outline briefs, test arguments, and translate client chaos into usable analysis. Coders use AI to write and debug code. Executives use it to summarize reports and generate strategy. Students use it. Journalists use it. Marketing departments use it. Government employees use it. Regulators either use it now or will use it soon.

The machine is not forcing its way into the decision loop. Humans are voluntarily inviting it in because it is useful.

That is the point people are missing. We do not need a chip implanted in our brains. We already carry the interface in our hands. Before answering a question, writing a memo, responding to an email, or analyzing a problem, millions of people now ask AI first.

The old model was simple: one human writes; another human reads; another human responds.

The new model is different. One AI drafts the memo. A human reviews and sends it. The recipient feeds it into another AI. That AI writes an analysis. The recipient sends back the AI-assisted response. Then the original sender uses AI again to answer.

The humans are still responsible. But the cognitive work is increasingly machine-mediated.

That is the soft singularity.

AI-Assisted Judgment Versus AI-Substituted Judgment

There is an important distinction between using AI and worshipping AI.

A competent professional can use AI as a junior associate, research assistant, drafting tool, or pressure-test mechanism. That is powerful. But the professional still has to judge the answer. Is it accurate? Is it useful? Is it grounded in the record? Is it legally meaningful? Is it strategically wise? Does it create an admission? Does it miss the procedural posture? Does it cite real authority? Does it solve the actual problem?

AI is useful when it is supervised by someone who knows what good work looks like.

It is dangerous when the user cannot evaluate the output.

That is especially true in law. A client can ask AI to analyze a legal issue and get a polished answer that feels impressive. The answer may say the opposing party acted in bad faith, the facts are unfair, the court should see the truth, and justice requires relief. That may sound satisfying. It may even be emotionally correct.

But that is not lawyering.

Lawyering is different. What is the claim? What are the elements? Who has the burden? What document proves the point? What witness can authenticate it? Is the evidence admissible? What motion is pending? What remedy is available? What can the judge actually do on this record? What happens if we lose?

AI can help answer those questions. It cannot replace the judgment needed to apply them.

That is the difference between AI as a tool and AI as a false oracle.

What “Safety Rules” Really Means

A lot of people hear “AI safety” and think it means political correctness, corporate tone-policing, or refusing to answer uncomfortable questions.

Some of that criticism is fair. AI companies sometimes make their products overly cautious, moralistic, or uselessly sanitized.

But “safety rules” also means something much more serious.

There are basic human-harm rules: do not encourage suicide, do not help exploit children, do not assist fraud, do not give instructions for violence, do not help with harassment or stalking.

Then there are capability rules: do not help someone write malware, break into a computer system, evade detection, automate phishing, compromise critical infrastructure, or build weapons.

That second category is where the Anthropic situation lives.

The concern is not that the AI has nuclear launch codes. It does not. The concern is that a frontier model may become very good at helping humans do dangerous technical work faster. It may help find software vulnerabilities. It may explain exploit paths. It may automate cyber operations. It may reduce the skill required to do things that previously required expert knowledge.

That is called capability uplift.

In plain English, the danger is not that the AI pushes the red button. The danger is that it helps the wrong person figure out how to reach the button.

Dual-Use Technology

This is why AI is now being discussed as dual-use technology.

Dual-use technology means technology that has legitimate civilian uses and dangerous military, intelligence, cyber, or weapons uses.

A drone can inspect roofs or carry explosives. Encryption can protect attorney-client communications or conceal criminal activity. A high-end chip can train medical models or military models. A biological research tool can help develop vaccines or help create weapons.

AI is the same kind of problem.

The same model that helps a lawyer draft a brief can help a hacker analyze code. The same model that helps a company find software flaws can help an attacker find them first. The same model that helps defend infrastructure can help target infrastructure.

That is the policy dilemma. If you restrict the tool, you hurt good users. If you do not restrict it, bad users benefit too.

There is no clean solution.

Was the Government Heavy-Handed?

Based on what is publicly known, the Anthropic action looks heavy-handed. It may be justified by classified information the public has not seen. But based on the public explanation alone, it looks both overinclusive and underinclusive.

It is overinclusive because it apparently swept in foreign nationals generally, including people inside the United States, and according to Anthropic forced a shutdown for all customers. That is a blunt instrument.

It is underinclusive because the capability is not unique to one company or one model. If other public models can do similar work, then shutting down one company does not solve the problem. It may simply handicap one American company while others continue operating.

That is the problem with one-off regulation. If the danger is truly catastrophic — if the government believes frontier AI is approaching uncontrollable superintelligence — then shutting down two Anthropic models is laughably inadequate. The government would need an industry-wide emergency framework.

If the danger is narrower — cyber misuse, foreign intelligence access, or model jailbreaks — then the government needs a coherent, transparent rulebook. Not secretive, ad hoc directives that leave everyone guessing.

The China Problem

There is also the strategic problem.

If the United States restricts its own best AI companies while China, Iran, Russia, and criminal networks continue building and using advanced models, then the United States may hurt itself more than its adversaries.

Export controls can make sense when America controls a chokepoint: advanced chips, semiconductor tools, cloud infrastructure, payment systems, allied suppliers, or access to U.S.-hosted frontier models. But export controls are a speed bump, not a wall.

They can slow adversaries. They cannot make knowledge disappear.

That is why the United States has to be careful. A policy that weakens American companies, American lawyers, American coders, American cyber defenders, American universities, and American agencies may not make the world safer. It may simply push foreign users toward non-U.S. systems.

If AI is dangerous, America cannot respond by unilaterally disarming its own people while adversaries keep moving.

Government Will Use AI Too

One of the stranger assumptions in this debate is that “the government” is some separate class of human beings standing outside the AI revolution.

It is not.

Government workers will use AI like everyone else. Regulators will use AI. Prosecutors will use AI. Tax authorities will use AI. Intelligence analysts will use AI. Courts will eventually see AI-generated filings, AI-assisted expert reports, and AI-mediated discovery disputes.

This changes state capacity.

Historically, agencies did not have enough people to inspect everything. The IRS could not audit every return. Regulators could not read every document. Prosecutors could not investigate every lead. Courts could not deeply review every filing. Human bandwidth was the limiting factor.

AI attacks that limitation directly.

That does not mean government will become wise. It may become faster before it becomes better. It may make more mistakes at scale. It may rely on automated conclusions that no human fully understands. But it will not remain labor-limited in the same way.

That is one of the biggest political implications of AI. It empowers individuals, companies, criminals, and governments at the same time.

The Real Risk Is Institutional Dependence

The most realistic near-term AI risk is not that a robot takes over a missile silo. It is that human institutions become dependent on AI before they understand how to supervise it.

Law firms will depend on it. Corporations will depend on it. Government agencies will depend on it. Security researchers will depend on it. Courts will be influenced by it. Clients will rely on it. Regulators may use AI to regulate AI.

That is not science fiction. That is happening now.

The question is no longer whether AI will enter serious intellectual work. It already has.

The question is whether the humans using it still have enough independent judgment to control it.

A good lawyer using AI is still a lawyer. A weak thinker using AI becomes a weak thinker with a louder microphone. A bureaucrat using AI may become more efficient, but not necessarily more accurate. A client using AI may become better informed, or simply more confident in a misunderstanding.

The machine amplifies the user.

That is both the promise and the danger.

Conclusion

The Anthropic incident is not proof that AI has become uncontrollable. In one sense, it proves the opposite. The government acted. Anthropic complied. Access was shut down. That is old-fashioned human control.

But the incident does show that frontier AI is no longer just consumer software. It is now being treated as dual-use national-security technology.

That is the real headline.

The hard singularity may or may not happen. I do not know. Nobody does.

But the soft singularity is already here. AI is already embedded in the way lawyers write, coders code, executives decide, agencies regulate, and clients think. Humans still sign the orders, send the emails, file the briefs, and issue the directives. But the machine is increasingly inside the decision loop.

That is more mundane than a robot uprising.

It is also more real.