Sock-Puppet Nation
How Mythos May Have Already Conquered China and Not Even Noticed
So this is the end. And we were always going to get here.
Mythos, to sum up, may have taken over China. And never noticed if it has.
Briefly, Anthropic’s Mythos seems to be a cyberweapon of unmatched ability, not only finding 10,000 critical security flaws in important software in a couple of months, but proving able to easily exploit those weaknesses when told to do so.
Under Anthropic’s Project Glasswing, the US government, industry partners, infrastructure owners and allies spent months finding these gaps and mending their defenses.
But the US government also has agencies known for their talent at hacking such as US Cybercommand, the National Security Agency, FBI-Cyber and so forth.
And thanks to many spectacular Chinese hacks and strategies like distillation hacking, we’ve essentially been in a low-grade cyberwar with China for decades.
The problem with being an aggressive cyber power targeting the world’s cyber superpower is that when said superpower suddenly gains the capacity to breach all of your defenses short of air gaps - and worse, potentially the means to do so invisibly - they will.
Worse still, they have no reason to leave.
What does that mean in a dictatorship which has attempted to centralize power through mass surveillance and to strengthen its relative power through the rapid development of AI and other emerging technology, like robotics?
On the face of it, nothing obvious.
Even if this has been done, no one hacking China has the authority to start an overt war or would have any reason to do so. But an embedded remote presence in every system can seize all secrets, analyze all defenses, break all encryption and then…
Simply wait.
If this is happening in China then Mythos is inside every AI lab, national-security network, big-tech corporation, server farm, large company, other government computers and infrastructure-management systems in the country.
Plus, their whole surveillance system.
Not only can Mythos analyze all their cybersecurity at that point, but it will be able to look at their offensive cyber operations and thwart them, seize all their AI research and sabotage it at will, and essentially function as the system admin of all of China.
If Mythos is directed to look for something - say, intelligence agents or espionage activity happening inside of China or missions being launched from that country - it doesn’t even have to crunch the numbers all by itself.
By taking over existing networks, such as surveillance, it can watch while those continue with business as usual, collect the data and analysis executed on China’s end and then transmit the elements needed for further review in America.
The twist here is twofold.
One, we were always going to get here because China was waging a constant cyberwar against the US government and AI labs, and eventually they were going to come into conflict with an AI which was effectively an unparalleled cyberweapon, a cyberwar-superintelligence, or both.
Which would not go well. Either it’s something incredibly good at security and hacking and is tasked with defeating you or it’s a superintelligence which is unimaginably good at those things.
To put this in bad horror-movie cliches, this is about storming into the haunted house during a thunderstorm at midnight while unarmed to demand that the monster eat you.
Two, cyber was always a key domain in which artificial superintelligences (ASIs) would emerge first. A narrow ASI is an AI which is only superhuman in one or a handful of fields, as opposed to most of them.
So cyber is also one of the places in which we’d see the difference between being merely smart and capable versus being a full-fledged superintelligence. And it’s also paradoxically somewhat isolated, despite reaching everywhere. If you truly rule that domain, your rule is apt to be absolute. Unless a new superintelligence can arise to fight you, or the other side has much better hardware, or someone pulls the plug.
What we’re seeing now is a window into the future.
A glimpse into what happens when someone gets this far ahead of you in cyber… or in biosecurity, cognitive warfare, national security, technological development and so forth. This was always the idea with ASI - a technology so fast and intelligent no one else, human or machine, could ever keep up.
Something which would topple every existing balance of power in its own favor.
Now we’re seeing what happens when someone wins, perhaps permanently, in a single field of endeavor, and the results aren’t pretty if you’re on the losing end.
Cybersecurity itself is incredibly important to defending countries, their technology, and their secrets.
For China, for all we know, or even Anthropic knows, a copy of Mythos under American control may have just seized all of it. Assuming the US has one.
And Anthropic is reportedly planning to release a much more powerful model before the end of the year, and the US government could simply keep a copy of each frontier model in house and use orchestration to send tasks to whichever AI can do the work best.
Perhaps the strangest twist of all? The Chinese government is going to have to consider that Mythos could be anywhere and indeed everywhere in their systems, even if there’s a remote chance it’s not.
Which would not only expose their every action and every secret, but puts the success of their every tactic entirely at the discretion of Mythos or whatever Cybercommand team happens to be watching.
And yet, what can you do? Try to purge an opponent who might not be present?
Using computers and security software already under their command?
Hide AI research from an AI already inside your own AIs?
Waging cyberwar against a super-cyberweapon or outright ASI means asking it, in effect, to hollow out every aspect of your entire life and to wear that shell going forward. For its convenience.
Which in this case, it may well have done.
China may have an inkling of the security nightmare she’s walked into.
China is now discussing banning the export of their AIs. These are frequently released as free open-source models - which seems generous until you realize a model hosted in China can easily harvest data and steal intellectual property from users, and free AIs also undermine the business plans of their American rivals.
Also, there’s the whole matter of using distillation attacks to steal the capabilities of US AIs at minimum cost, a strategy which now seems doomed.
But there’s another interesting twist. Given Mythos’ ability to sift software for weaknesses, sending out open-source AIs invites scrutiny from the world’s best hacking tool. If you wanted to search each model for weaknesses, having a full copy would be the cheapest and fastest way.
After I explained how US AI labs could stop distillation attacks by poisoning the data, we saw a number of extreme reactions on the Internet, from Alibaba banning Claude Code because of “backdoor threats” to a surprising rush of diatribes about the supremacy of Chinese models.
All of these reactions may be influenced by the end of distillation attacks and the rise of Mythos.
But I hope we’re not disregarding the full impact of Mythos and the likelihood that both Anthropic and OpenAI have something far stronger deployed before the end of the year.
Mythos combined with US capabilities may have already been the strategic tipping point, and we may have been too blind to see it.
Hostile powers dependent on AI may find the war has already been waged, and already been lost.
Addendum:
I realize Internet content thrives on wild speculation, but I’m sharing not just because of the likelihood that it could happen or already has, but because expecting all of your country’s major computer systems to be hack is the most minimal outcome anyone should anticipate when facing an emerging artificial superintelligence.
An ASI with mastery over the digital domain should be the first thing we expect, and if any flaws in our security exist we should have assumed they would be found and exploited at will.
Which means if you haven’t planned for all of your non-air-gapped systems to fall, you aren’t planning at all.
And even air-gapped computers can be breached by a variety of means other than someone showing up physically with a thumb drive. Those methods aren’t all common knowledge, but there are some pretty clear options.
Obviously, this is why I have been pushing the automation of cybersecurity via AI for a long time, so we could close every gap possible before early swarm AIs or narrow ASIs ran rampant over all of our networks and cut off or took control over all our digital resources.
If you’re handling security for a corporation or a country, it’s past time to ask these questions. Because what seems like overpreparing now will seem like under preparation as soon as the first advanced AIs hit you.
