Technology

Earlier this year, Antrhopic announced a large language model (LLM) called Mythos, a model they claimed to be so powerful for finding software exploits that it was too dangerous to release to the general public¹²³. Instead, they provided a limited release of this model to select organizations as part of Project Glasswing⁴. Then, last week, Antrhopic released a version of this model called Fabel⁵. Current subscribers were allowed an entire three days to work with this model before the US federal government issued an unpublished export control order for Anthropic, citing concerns of model jail breaking and weaponization, to restrict access to the model from any foreign nationals living in America or abroad⁶. Anthropic denied the government’s finding, but complied with the order by revoking access to the model for all customers⁷. There has been a massive roller coaster of press over this “frontier model,” much of it seeming to be media-hype and exaggerated claims. However, I question if this latest intervention by the US government is really about this coding model. The names of the products, the timing and the general story surrounding technology leads me to believe this is a narrative propaganda tool being used to increase government regulation, cripple competition and push for increased government surveillance.

The Myth and the Magic

Anthropic’s models have names like Haiku, Opus and Sonnet, all which seem related to songs or poetry. But this most recent model started as Mythos, and the released version was Fable. Both of these names invoke ideas of fantasy or mystical fictions. As a developer, my peers and I got to play with Fabel for all of three days, before it disappeared over a weekend. Think about that: software engineers got to use it for work, and by the weekend it was no longer available, even on our personal accounts.

I used Fable pretty heavily on some complex features while it was available. I was impressed by some of its planning outputs and did find it somewhat cleaner than previous models. My friends and colleagues have given me various types of feedback, some amazed by it while others seeing it as a disappointment or incremental improvement.

The Call for Regulation

Earlier in June, Anthropic had called for a “global pause” on further AI development due to risks of “self-improvement⁸.” So, they called for a pause and then turned around and made a new product release? What I find most concerning is a post from Anthropic’s founder calling for government regulation⁹.

It is time to go beyond transparency to more serious and binding regulation of AI. I believe the best analogy, at least at the current stage of the exponential, is to cars, airplanes, or drugs—powerful technologies essential to the modern economy, but capable of killing large numbers of people if designed or operated poorly. I therefore believe we should model AI regulation on agencies like the Federal Aviation Administration (FAA). Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety. I am grateful to see the Trump administration’s Executive Order move incrementally towards a greater role for government in AI, though Anthropic’s proposal recommends even further action. -Dario Modei⁹

Whenever a company within an industry calls for more regulation, it’s rarely in the public’s interest. It’s usually a means of preventing competitors from entering their market. What would regulation even look like in this case? Would developers be required to submit any software that multiplies matrices together to the United States government? Will linear algebra software fall under export regulations? Will there be more cases similar to Roman Storm and Tornado Cash, where simply writing open source software could get someone arrested for evading sanctions?

People may say, why didn’t you just disable access to foreign nationals? Why did you disable access to everybody? And this comes back to an issue we talk about a lot … How are you going to keep foreign nationals from using it? When I sign up, all I need to provide is an email and my credit card information … That’s all I need to sign up for one of these services … but how do they know if I’m a foreign national, US citizen or anything else? Maybe my ID will show that. Maybe my ID is not enough. Maybe I got to show my passport. Maybe I show my ID, my passport, and my birth certificate to show proof. You see where we’re going with this? This is moving the internet in the direction of you need to show your passport, your birth certificate, and your ID to sign up for a service … And to do that would require that you put all this infrastructure in place … I’ve spoken about a lot of the age verification laws that have been popping up recently and why I’m annoyed by them … This is bad. I don’t like this. And this kind of speed-runs you further in that direction. -Louis Rossmann¹⁰.

The entire story with the release and redaction of Fable, followed by calls for regulation, reeks horribly of surveillance-state propaganda. It creates a narrative of fear, which could provoke citizens to call for protection in a way that stifles innovation and pushes us further into surveillance-state policies.

Scary Security Stories

Anthropic claimed that their Mythos model detected 23,000 potential vulnerabilities across a thousand open source projects¹¹¹². This included a “27-year-old bug in OpenBSD¹³,” which was patched¹⁴, although I can’t find any official statements on it from OpenBSD developers regarding the severity. FFmpeg did thank Anthropic¹⁵ for patching a vulnerability in their H264 decoder¹⁶. The developer of cURL reported that a scan with Mythos resulted in it finding a single vulnerability¹⁷, while also stating that Mythos was “an amazingly successful marketing stunt for sure¹⁸.”

The Fable of Defying Trump

Anthropic has allegedly been at odds with the Trump administration. They initially tried to take the ethical high ground, refusing to allow their models to be used for fully autonomous weapons. This caused the Defense Department to label Anthropic a supply-chain risk¹⁹, which then resulted in Anthropic suing the federal government²⁰. This moral grandstanding was short lived as shortly afterward, people discovered that Anthropic was experimenting with displaying different features to customers randomly on their Claude Pro product, in order to experiment with seeing how much people would pay for additional feature sets²¹²².

So, What Just Happened?

One of the supposed jailbreaks came from “Pliny the Liberator,” a person who continually tries to prompt inject every newly released model²³. Another person was able to bypass the “review this code for security issues” guardrails by using the prompt “fix this code²⁴²⁵.” TechCrunch argued that the issues should have never triggered export control violations and that the government ban was never about jailbreaks²⁶.

Going by Dario’s statements on practically begging for regulation⁹, it’s likely this entire fiasco is about regulatory capture. As of writing, AI is profitable for absolutely nobody except Nvidia²⁷. Theories have also emerged that the LLMs are so unprofitable that the dangers of Mythos were intentionally exaggerated as an attempt to force the US government to use eminent domain to purchase Anthropic at fair market value²⁵.

The Future of A.I.

I’d like to think many people are weary of the push for Large Language Models into every aspect of our lives²⁸. Still, I cannot deny that, for software engineers and other professionals, they are incredibly powerful tools for rapid prototyping or creating workflows for specific, well defined tasks. There is a capture taking place, among the engineering-class of society, that wants to alleviate the painful details we often have to exercise for meticulous tasks. This might be all well-and-good for people who write billing or HR software. However, I do wonder if at some point we will see a bridge collapse, and the blame will be placed on an inference model that incorrectly text-completed (or as some say “hallucinated”) the tensile strength needed to hold the expected load; something that might have been caught by a human engineer vested in the safety of their product and the real humans walking across it.

In the recent Anthropic fiasco, I do think we are watching the complexities of an early onset market collapse around the LLM industry. The names of Mythos and Fable invoke a fantastical storytelling narrative, which I think were intentional. Anthropic has political motivations combined with exponential spending. Their fearmongering and calls for regulation could be a strategy for regulatory capture, increasing the costs for competitors to enter their product space. We are speeding forward toward a tipping point, because the subscription rates people pay likely don’t remotely begin to pay for data center costs²⁹³⁰.

We are being sold a myth right now³¹. It’s right there in the name. The entire situation with the technology sector today feels way too close to the 2008 financial collapse, combined with the 2001 dot-com bubble, and then multiplied by an order-of-magnitude coefficient large enough to collapse the global economy.

We don’t know if Anthropic is acting independently, or if they are actively colluding with a government (or multi-government) narrative. The names, timing and media stories makes me skeptical of everything surrounding the entire Mythos/Fable shipwreck. We are literally living through the mythology of a software engineering fable, and I’m very weary of where this all leads.