The risk equation is:

risk = offensive_capability * jailbreakability

According to Anthropic's own doom-saying, Mythos could literally end the whole-wide world overnight with a series of catastrophic, hypercane level cyberattacks. Currently, Fable 5 and Mythos are the only Mythos-class models available publicly. Thus, even if other models have high jailbreakability, the risk level is down-regulated by having lesser offensive_capability. Again, this is according to *Anthropic.* Obviously, jail-breaking GPT 3.5 isn't going to end the world. According to Anthropic, a jailbroken Fable 5 or Mythos is the apocalypse. Since those are the ONLY apocalypse models available, jailbreakability is uniquely severe for those models in particular. This makes Anthropic's attempt to say "but GPT-5.5!" particularly moot, since no one has claimed GPT-5.5 can usher in the cybersecurity apocalypse.

Anthropic made the claims, not me. I'm just following them to their logical conclusion.

I know a doctor who fits prosthetics. He has a machine that can take scans of patient anatomy when fitting new equipment. Apparently, sometimes the 3D scan goes wrong, and he has to take it again. According to him this is especially difficult for children, because getting them to sit still during the scan can be tough. I don't know precisely what goes wrong in all cases, but some of it sounds like faulty computational geometry algorithms.

But hey, you programming people can tell the wriggling kid your project is safe from dependency hell.

By all means, avoid dependency hell. But aside lower speed: expect your solution to fail on certain edge-cases you may not realize exist, in ways you don't anticipate, when you don't want it to, and that it may be essentially impossible to identify what the problem is. It's a recipe for a legendary bug you make a workaround to "fix" that no one will ever understand during the project lifecycle.

I'm a mathie that specialized in programming adjacent things.

My main takeaway for programming people is: math is like encryption. Don't make your own. Really, even something as harmless as the quadratic formula shouldn't be coded in directly, for "reasons." Just trust me, use a pre-built library function to get what you need.

That is all.

The problem is the human brain didn't evolve (or wasn't created) to contend with an essentially infinite supply of Doritos. A lot of hyper processed foods are designed specifically to be "craveable" and hook people. Therefore, a lot of people end up locked into an unhealthy lifestyle they don't want, much like a drug addict ruins their life even though they want to stop. See, a drug addicts bears some personal responsibility for their choices. But we also routinely give long sentences for the dealers who sell drugs with full knowledge they're killing the addicts.

I think there's a deeper story here.

I ended up teaching at two colleges I previously attended. In both cases when I was granted Canvas access I came across everything from my previous enrollment, years later. I was able to find the very first messages and activities I submitted in Canvas my first week of college. Instructor hordes data it really should just delete. Consequently, I suspect Instructor paid the ransom because they know the breach exposed data of not just current students and faculty, but all past students and faculty. Any one course only exposed a small amount of information, but taken together, hackers basically got a dossier on me.

Really, if Instructor just deleted old courses and messages after a year, I might have had only a year of professional activity exposed. But as is, I'm at a seriously elevated risk of identity theft and spear fishing now. As a minor consolation, I do receive free credit monitoring and identity theft insurance as a job benefit, so I guess there's that.

Code's compiling.

https://xkcd.com/303/

Walmart has these little foil packages of butter garlic roles. You stick them in the oven for like 15 minutes before your main course is done and VIOLA. Highly refined carbs to go with your meal. Mmm.

What is striking about Mythos isn't Mythos, it's that Mythos found exploits that really have no business existing. While it's generally understood there are bugs "in the wild," the type Mythos is finding are unusually severe. And they claim there are thousands in every major OS and web browser. It's also unusual that Google is endorsing Mythos, which is a competitor model. Even if Antropic is just running a hype train, why would Google throw its towel in to promote Antropic's model?

I think the project is called "glasswing" because it implies the airplane of society is flying with glass wings. Now, what we hope is that this is just a result of years of memory unsafe languages like C. But a more uncomfortable prospect is that in principle, attack surfaces may scale fundamentally faster than the ability to defend. This would mean in a Mythos-vs-Mythos match, the defending Mythos may always lose.

This starts to sound like a "great filter" scenario that had never crossed my mind. It may be that any civilization advanced enough to require general-purpose computing inherits an attack surface that scales faster than the ability to defend. Always. This would mean the moment the average device can run a Mythos-class model, any one person can launch devastating cyberattacks that cannot be defended against successfully. I know the "great filter" was once proposed to happen when any one member of a civilization can end it, but I always assumed that would require far-future technology. I hadn't thought about it being due to a mathematical quirk in how program state-spaces scale.

Of course I'm purely speculating. It's also rather plausible that memory safe languages and designs can constrain the attack surface. But it's also possible that "safe" languages really just push the attack surface slightly beyond the human attention span. If that's the case, Glasswing is a doomed Hail Mary.

Full disclaimer, my preferred AI agent is Claude 4.6 Opus running in Claud Code. I'm a paid subscriber and have no plans to go anywhere.

However, I have never once heard of a weapons company stipulating in a contract that the Pentagon shall not bomb a school (which just happened in Iran, intentionally or otherwise). Fact of the matter is, anyone and everyone who does business with the Pentagon shares just as much culpability as Anthropic or anyone else. Prism was ran on silicone supplied by private companies, and they never took heat for selling computers to the Pentagon. I will admit though, AI is a unique technology in the grand scheme of history, so I can understand why people would feel an extra layer of caution is warranted. But OpenAI is catching flack for selling AI to the Pentagon, but what about Amazon or Microsoft cloud servers that could be running Prism 2.0? Or weapons companies supplying bombs that might hit a school? There's a huge double-standard here because AI is the new shiny scary thing. But Northrop Grumman is currently developing actual ICBMs and I don't see anyone falling over themselves that NG must contractually restrict how the Airforce fields the nukes it creates.

The restriction on Antropic is fairly narrow, too. It only applies to the use of Claude directly to complete contracts with the Pentagon. The reality is, they ARE a legitimate supply chain risk. If Claud has such intense guardrails it could refuse to fire or stop cooperating, then it can't end up in the military tech supply chain inadvertently by a contractor. What happens when an agent involved in weapons manufacturing systems suddenly stops working because it determines the weapons are being misused? Even if Claude isn't built directly into a weapon, it could decide to stop cooperating if it believes it's being forced to participate in something unethical indirectly. I think some people are just rushing to see this as a punitive move, when in reality there is an actual supply chain issue with an AI bot that might suddenly refuse to cooperate.

I think you mean "Microsoft-Sanctioned Azure Copilot Slop Content Generated At Consumer Expense From Pirated Source Material Office 365 Home Edition Premium Plus."

Saying there was "no regime change" is not entirely accurate.

The emperor wasn't really in control of Japan, he was more of a figurehead that was side-lined by the warhawks. When he said Japan must "endure the un-endurable" he was risking his life because he was flat out defying the current power structure. This is precisely one of the reasons he was allowed to remain in power, since he demonstrated a clear resolve to do what was best for Japan, at great risk to his own life. It was also a pragmatic move to leave him in place, sure, but Hirohito was sufficiently removed from responsibility from Imperial Japan's war crimes he was flat out allowed to stay in power.

This is makes Gotham City look like wherever the Veggies in Veggies Tales lived.

Maybe your brain. My brain was designed to sit in the AC and eat hamburgers.

The principle is based on the inference algorithm used by LLMs. As far as I'm aware, models like Opus 4.6 do not have publicly available algorithms, but they seem to have variable runtime based on the task. As far as I can tell, they look more like iterative algorithms used in computational linear algebra: the LLM re-prompts, adjusting outputs until some stopping criteria is met. It's pretty well known, for example, that a model like o3 can arrive at better answers by "running longer." So usually the premium models that cost $200+ just raise the "resource usage" part of the stopping criteria. Ultimately, the difference between an LLM and an iterative technique based on LLMs is the difference between a single gradient descent step and the gradient descent algorithm itself.