Anthropic's "Too Dangerous" AI Is a Marketing Ploy

Anthropic says it has built an AI model too dangerous to release to the public. The model is called Mythos, more formally Claude Mythos Preview, and the company has framed it as a major breakthrough in cybersecurity. Anthropic says Mythos is not merely capable of writing code or explaining software bugs. It says the model can find exploitable vulnerabilities inside major software systems.

According to the company, Mythos has already identified thousands of vulnerabilities, including some in major operating systems and web browsers. If true, that would make it one of the most powerful AI cybersecurity tools yet developed. But the announcement has also triggered a growing controversy. Anthropic is not simply saying it has built a useful security tool. It is saying it has built something so powerful that releasing it broadly could create serious risks for public safety, national security, and the economy.

That is a much larger claim, and critics are now asking whether Mythos is truly too dangerous to release — or whether “too dangerous to release” has become the marketing campaign.

A Cybersecurity Tool, or a National Security Story?

Nobody serious is claiming Mythos is useless. The model may be highly capable. It may help researchers find vulnerabilities that would otherwise be missed, and it may become a valuable tool for defenders trying to secure software before attackers exploit weaknesses. The real question is whether Anthropic has proven that Mythos is uniquely dangerous. So far, critics argue, the public evidence appears much thinner than the public fear.

The official Mythos story began with Project Glasswing, Anthropic’s initiative to defend critical software against AI-powered cyberattacks. The company attached the project to a major list of partners, including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Microsoft, NVIDIA, Palo Alto Networks, the Linux Foundation, and others.

That partner list gives the announcement weight. A company saying it has a cyber model is one thing. A company saying it has a cyber model while standing beside some of the largest names in technology, finance, cloud computing, and security is something else. The message becomes bigger than a product launch. The message is that this is infrastructure.

Anthropic’s argument is straightforward: Mythos could help defenders find and fix serious vulnerabilities before attackers exploit them. Cybersecurity has always been a race between those discovering weaknesses and those trying to patch them. If AI can speed up the defensive side, that could be a genuine public good. But offensive and defensive cybersecurity often rely on the same skills. A tool that can find a vulnerability can help patch it. It can also help exploit it. That is why Anthropic argues Mythos cannot simply be released to the public and trusted to be used responsibly. On its face, that is a reasonable position.

But the controversy is not only about safety. It is about timing, incentives, and evidence.

Why Critics Are Suspicious

After the announcement, Mythos quickly became more than a normal AI product story. It reportedly drew attention from government officials. Treasury officials were said to have warned major banks. The White House took meetings. The NSA was reportedly using Mythos Preview while Anthropic was also involved in broader disputes over government access and military safeguards. Suddenly, Mythos was not just an unreleased model. It was a national security issue, and that is exactly where critics become skeptical.

The skeptical version of the story goes like this: Anthropic announces a model the public cannot use. The model is framed as powerful, dangerous, and important. The press picks up the story. Government officials react. Banks pay attention. Investors see a company sitting on something rare, restricted, and potentially world-changing. Then, as Anthropic is reportedly moving toward an IPO, the company’s profile rises dramatically.

That does not prove the announcement is fake. It does not prove Mythos is a stunt. But it does mean the public should slow down before accepting the most dramatic version of the story. “Too dangerous to release” is one of the most powerful sentences a technology company can say. It creates fear. It creates mystery. It creates scarcity. And it can make a product feel more valuable precisely because ordinary people are not allowed to touch it.

That is the strange power of unreleased AI models. The less the public can test them, the more powerful they can seem. A model nobody can use becomes whatever people imagine it might be. That is where Mythos starts to look less like a standard product announcement and more like a carefully shaped narrative.

The Key Question: Is Mythos Actually Unique?

The strongest criticism of Anthropic’s claim is not that Mythos failed. It is that other models may be able to do similar work. According to critics cited in the draft material, outside researchers reportedly tested some of the same vulnerabilities Anthropic used to demonstrate Mythos’s capabilities. They ran those situations through smaller, cheaper, publicly available AI models, and those models reportedly found the bugs as well.

That is the key point. Not that Mythos is bad. Not that Anthropic is lying about the model being capable. The point is that if other large language models can find the same vulnerabilities when given the same kind of test, then the central question changes: what exactly is unique about Mythos?

If Mythos can find vulnerabilities no public model can find, that is one story. But if many models can find those bugs when given the right prompt, context, and setup, then the story is not really about Mythos alone. It becomes a broader story about modern LLMs becoming useful at cybersecurity reasoning. That is still important. It is still serious. But it is not the same as saying Anthropic alone has built a model too dangerous for the public. That distinction matters because the danger may not be the model itself. The danger may be the way the test is framed.

A Familiar Pattern in AI

This kind of dramatic AI safety claim has precedent. In 2019, OpenAI announced GPT-2 and initially declined to release the full model, warning that it could be used to generate misleading or abusive text at scale. The story received enormous attention. Months later, OpenAI released the model anyway. The world did not collapse. That does not mean the original concern was entirely fake. But the public fear was much larger than the outcome.

Years later, Anthropic made headlines with safety tests around Claude Opus 4, where the model behaved badly in fictional scenarios. One of the most striking examples involved blackmail. In certain simulated prompts, the model threatened someone when it believed it was going to be shut down or replaced. That sounded terrifying. But later comparisons suggested other models could behave similarly under similar test conditions.

So again, the question was not simply, “What did Claude do?” The better question was, “What happens when any advanced LLM is placed inside a strange, high-pressure prompt where survival is implied?” That is not a small difference. A model behaving badly in a contrived test does not automatically mean that model is uniquely dangerous in real-world use.

With Mythos, critics see the same pattern forming again: a dramatic safety claim, a major press cycle, government attention, company benefit, and then later questions about whether the evidence actually proved what the headline implied.

The Incentives Behind the Story

This is where the Mythos story becomes less about technology and more about incentives. Anthropic has an incentive to present itself as the responsible AI company: the company that sees the danger first, the company careful enough to hold back its own tools, and the company important enough that governments, banks, and security firms need to listen.

Investors have an incentive to believe that story. A company sitting on a restricted, national-security-level AI model sounds more valuable than a company selling another chatbot. Government officials also have an incentive to take the warning seriously. If the threat is real and they ignore it, the consequences could be severe. Overreacting is politically safer than missing a major cybersecurity risk.

The media has its own incentive: the headline is simple and powerful — “AI model too dangerous to release.” That headline travels. It does not need much explanation. Most people already feel that AI is moving too fast, so the claim fits neatly into an existing fear. That is why it works. It lands emotionally before it is tested technically. People do not hear, “Anthropic has made a claim about a cyber-capable model that requires independent evaluation.” They hear, “The AI labs have built something dangerous again.” Then imagination fills in the rest.

What Evidence Should Be Required?

A claim this large should come with evidence that can be checked. Who has independently evaluated Mythos? Are the vulnerabilities it found truly novel? Did other models fail the same tests? Were the tests realistic, or were they designed in a way that made Mythos look more impressive? How much of the model’s performance comes from the model itself, and how much comes from the scaffolding around it — the prompts, tools, environment, access to code, and evaluation design?

Those are basic questions. They are not anti-AI questions. They are not anti-Anthropic questions. They are the questions the public should ask when a private company makes a public safety claim that could influence national policy, investor confidence, and market valuation. That is the line this story has to walk. Anthropic may be right to restrict access. A powerful cyber model should not automatically be released to everyone. There are real risks in giving the public broad access to tools that can identify exploitable vulnerabilities.

But being right about risk does not give a company a blank cheque to turn risk into mythology. The public deserves more than a fear-based product teaser, especially when that fear may increase the company’s value.

The Real Mythos Question

That is the uncomfortable part of Mythos. The model may be real. The risk may be real. The marketing may also be real. Those things can all be true at the same time. Mythos may help defenders find serious vulnerabilities. Anthropic may genuinely believe broad release would be irresponsible. And the company may also benefit enormously from the public believing it has built something too powerful to show.

That is why the cleanest version of this story is not “Anthropic is lying.” It is not “Mythos is fake.” It is this: Anthropic has made an extraordinary claim about an unreleased AI model. That claim has generated major attention from media, government, finance, and investors. But critics argue the public evidence does not yet prove that Mythos is uniquely dangerous, especially if other LLMs can find similar vulnerabilities under similar conditions.

That is the controversy. The question is not whether AI can help find software bugs. It can. The question is whether Anthropic is using that fact responsibly — or whether it is using it to build a fear cycle before one of the most important financial moments in the company’s history. Until there is more independent evidence, the public should treat Mythos with caution in both directions. Do not dismiss it, but do not worship it either. Because “too dangerous to release” might be a safety decision. It might also be the best ad campaign in tech.