Too Dangerous to Release, or Too Profitable to Share?

Watch the framing. That is where the real work happens. Not in the model, not in the benchmarks, not in the partner announcements. The framing is the product. Everything else is packaging.

Anthropic told the world that Claude Mythos Preview is too dangerous for public release. Read that sentence again. A company that makes money selling access to AI models told the world that its newest, most powerful model is so dangerous that only its biggest paying customers should use it. And the press covered this as a story about safety.

“

We do not plan to make Claude Mythos Preview generally available due to its cybersecurity capabilities. -- Newton Cheng, Anthropic

The rhetorical move is elegant. Step one: build the most powerful AI in history. Step two: declare it too dangerous for the public. Step three: give exclusive access to twelve of the largest corporations on Earth, including your own investors and cloud hosting partners. Step four: frame the entire operation as selfless restraint. Step five: announce 30 billion in revenue on the same day.

In any other industry, this would be called what it is: a premium product launch with restricted distribution. Luxury brands have used this model for centuries. Scarcity creates value. Exclusivity creates demand. The only innovation here is that Anthropic replaced "exclusive" with "responsible" and "premium" with "safe."

“

Frontier AI capabilities are likely to advance substantially over just the next few months. It will not be long before such capabilities proliferate. -- Newton Cheng, Anthropic

Biased Bipartisans

Sponsored

Real-Time, Evidence-Based News Reports

Unlimited access to your personalized investigative reporter agent, sourcing real-time and verified reports on any topic. Your personalized news feed starts here.

Create Free Account

The language is worth dissecting. Newton Cheng, Anthropic Frontier Red Team Cyber Lead, told VentureBeat: "We do not plan to make Claude Mythos Preview generally available due to its cybersecurity capabilities." Note the passive construction. Not "we choose to restrict access." Not "we decided that profit margins are higher with selective distribution." The framing removes agency. The model is dangerous. The restriction just happens. Anthropic is not making a business decision. Anthropic is responding to an objective reality.

Now watch the pivot. In the same interview, Cheng acknowledged that similar capabilities will proliferate within months. Months. If adversaries will have equivalent tools within months, the "too dangerous" framing dissolves. You are not protecting the world from a unique threat. You are giving your customers a temporary advantage and calling it national security. The window of exclusive access maps precisely to the window of competitive advantage.

The security lapses are the gift that keeps giving. Anthropic left 3000 internal assets in a publicly searchable data store. An npm error exposed 512,000 lines of source code. The company that cannot secure a blog CMS and a package manager now claims the moral authority to decide which organizations are trustworthy enough to access a world-changing capability. When confronted, the response was that these were human errors in publishing tooling. Translation: the errors that matter are never our errors. The errors that count are the ones other people might make with our model.

Biased Bipartisans

Sponsored

Think Further on BIPI.

Where seeking the truth is a journey, not a destination.

Learn more

The 100 million in usage credits deserves the full treatment. That number anchored every headline. "Anthropic commits 100 million to cybersecurity." Impressive. Until you compare it to the 30 billion in annualized revenue disclosed the same day. The ratio is one third of one percent. The four million for open-source organizations is one one-hundredth of one percent. These are not commitments. These are marketing expenses classified as philanthropy.

Here is the move that nobody in the press caught. Anthropic proposed that an independent third-party body should eventually oversee AI cybersecurity projects. Eventually. Not now. Not before they lock in their market position. Not before the IPO. Eventually, after the competitive advantage has been extracted, after the revenue has been booked, after the partnerships have solidified, then maybe an independent body should take over. The proposal is not governance. It is an exit strategy dressed as institutional design.

The founding myth completes the picture. Anthropic was created by people who left OpenAI because they thought OpenAI moved too fast and cared too little about safety. That origin story is the most valuable asset the company owns. It is the reason every restriction gets covered as responsibility rather than strategy. It is the reason "too dangerous to release" reads as moral courage rather than market positioning. The founding story does the work that the product itself cannot do: it makes controlled distribution look like sacrifice.

None of this means the capabilities are fake. The vulnerabilities are real. The benchmarks are real. The 27-year-old OpenBSD bug is real. The question was never whether the model works. The question is whether the language used to describe its deployment is honest. And on that question, the evidence is clear. Anthropic built something powerful, restricted access in ways that benefit its commercial partners, committed a fraction of a percent of revenue to mitigation, and described the entire operation as an act of conscience. The audience applauded. That is what good framing does.