There are two groups that seem absolutely thrilled by the recent revelation that Anthropic is working on a top-secret AI model: the Department of Defense and Anthropic’s fundraising team.
On Thursday, Fortune reported that Anthropic left a slew of not-yet-public information accessible on its website, leading to the reveal of a supposedly super-powerful, yet-to-be-released model called Claude Mythos. The model—which is supposedly “by far the most powerful AI model we’ve ever developed,” per the text of a leaked, unpublished blog post from Anthropic—blows away benchmarks set by the company’s current public model, Claude Opus 4.6. The company confirmed the model’s existence to Fortune and called it “a step change and the most capable model we’ve built to date.”
The company also basically said it’s holding back the model because it might actually be too dangerous to release to the public at this point. Per Fortune, Anthropic said the model is “currently far ahead of any other AI model in cyber capabilities,” and presents unprecedented cybersecurity risks that the company seems to want to get out in front of.
That revelation has the Pentagon doing victory laps, even though it never gave any indication of having concerns about this type of thing before. The Department of Defense is still carrying a grudge against Anthropic after the company refused to agree to allow its model to be used for domestic surveillance or fully autonomous military weapons. And while it seems the Pentagon is losing its legal case—a ruling handed down from a judge Thursday temporarily blocked the DoD from labeling Anthropic a security risk—it’s still looking to win in the court of public opinion.
Following the reveal of Claude Mythos, Under Secretary of War and chief Anthropic antagonist Emil Michael posted, “Umm…hello? Is it not clear yet that we have a problem here?” Michael has been hammering Anthropic for weeks now, calling the company’s CEO Dario Amodei a “liar” with a “god complex” who wants to “personally control the US military,” and now he’s treating the leak as proof that Anthropic can’t be trusted.
Michael’s position is almost certainly not coming from a place of genuine concern. Anthropic was given the green light to handle classified material from the federal government, and most of the military applications for the model are operating within secure environments (or they are supposed to be). The fight has been about the Pentagon wanting to use Anthropic’s models even more than the company wants, which doesn’t seem like the thing you’d fight for if you thought the models presented security risks.
Michael has significant financial ties to other AI firms that are competitors of Anthropic, so it’s hard to take his view of the matter at face value. Also, this is the same Pentagon that added a journalist to a Signal group chat where war plans were being shared, so let’s not pretend like their security protocols are airtight.
Regardless, if we take the leak at face value, it seems like Anthropic is sitting on a big release, though who knows when or if we’ll ever see it. It’s reportedly very expensive to train and operate—so expensive, in fact, that the company seemingly isn’t sure how to deploy it at this point.
There’s little reason to believe that Anthropic left this information out in the open with the intention of someone finding it. Frankly, it’s a pretty embarrassing oversight to leave your content management system accessible without requiring a password, which is how this information all became public. But it’s also hard to ignore the fact that this whole situation plays right into the classic AI company playbook of talking up the dangers of a model to highlight how powerful and capable it is.
So, it’s fitting that around the same time as the leak about the extremely powerful model that blows everyone else away, The Information reported that Anthropic was eyeing going public later this year. That secret, super-powerful model does seem like the kind of thing Anthropic would like potential investors to know about ahead of an IPO.
Read More:
Kash Patel’s Email Account Hacked, Iran-Linked Group Takes Credit
Read the full article here
