Every AI Is 'Too Dangerous to Release' Until It Isn't

In February 2019, OpenAI announced it had created something dangerous. GPT-2, a text-generating AI, was so good at writing convincing prose that the company said it could not release the full version to the public.

Due to our concerns about malicious applications of the technology, we are not releasing the trained model.

— OpenAI, February 2019

The announcement generated headlines around the world. An AI too dangerous for public release! Scientists scared of their own creation! The coverage was extensive and, more importantly, free.

Nine months later, OpenAI released the full model anyway. The feared wave of AI-generated disinformation did not materialise. GPT-2 turned out to be... fine.

The Pattern Repeats

This week, internal documents leaked showing that Anthropic's next-generation model, Mythos, scored so high on cybersecurity and hacking benchmarks that the company is afraid to release it. The framing is familiar: too powerful, too dangerous, must proceed with caution.

One Reddit commenter put it bluntly: 'From Anthropic's perspective, this is just marketing. It's the same as OpenAI when they said our AI is too dangerous.'

The pattern has repeated with every major AI release since GPT-2:

• GPT-3 (2020): Initially restricted, then opened widely
• GPT-4 (2023): Delayed for 'safety testing', released six months later
• Claude 2 (2023): Positioned as 'safer' alternative, released on schedule
• Claude Opus (2024): 'Constitutional AI' framing, still released
• Mythos (2026): 'Too dangerous', leak generates global headlines

Every single model eventually reached the public. and not one has stayed locked away.

The Leak as Launch Event

Consider the mechanics of the Mythos 'leak.' Internal documents appeared online, were quickly reported by major tech outlets, and generated exactly the kind of coverage that launch events are designed to produce , without the expense of an actual launch.

Anthropic gets to position itself as cautious and responsible. The model gets positioned as powerful and desirable. And when Mythos eventually releases (as every model does) the company can point to this period of deliberation as evidence of their safety-first approach.

AI labs probably aren't lying about capabilities. GPT-5 really is more powerful than GPT-4. Mythos probably does score higher on hacking benchmarks than Claude Opus. The question is whether 'too dangerous to release' is a genuine safety threshold or a marketing position.

The Business Case for Danger

From a business perspective, the 'dangerous AI' narrative serves multiple purposes:

First, it generates press because announcing a new model gets coverage. Announcing a model so powerful you're scared to release it gets significantly more coverage.

Second, it positions the company as a responsible actor in a largely unregulated industry. When regulators come asking questions, the company can point to its history of caution.

Third, it creates scarcity and desire. A model you can't access is more desirable than one you can. When access eventually opens, users rush in.

What Would Real Caution Look Like?

If an AI lab genuinely believed a model was too dangerous to release, what would that look like? Probably not a leak followed by extensive press coverage of capabilities.

It might look like: not building the model in the first place. Or building it and never announcing its existence. Or releasing it only to vetted researchers under strict controls that don't expire.

Instead, we get a predictable cycle: dramatic announcement, safety concerns, public debate, eventual release, move to next model. The 'dangerous' AI joins the consumer product line, and the next model becomes even more dangerous.

The Cynical Reading

Here's the uncomfortable possibility: AI safety concerns, as articulated by the labs themselves, may function primarily as marketing. They create urgency, imply capability, and generate headlines.

AI risks are probably real. But the companies raising those concerns are also the ones benefiting most from the attention.

When Anthropic's CEO warns about existential risk while raising $30 billion at a $380 billion valuation, you have to ask: is this person sounding the alarm, or ringing the cash register?

The answer, most likely, is both. And that's exactly the problem.