Anthropic trained a model called Mythos. They did not train it to hack things. They trained it to be good at code. But as a side effect of being exceptionally good at code, it became exceptionally good at finding security vulnerabilities in software that humans have missed for decades.

The model found exploits in OpenBSD, a famously secure operating system. It found vulnerabilities in FFmpeg, a foundational piece of open source video infrastructure. It chains together three, four, sometimes five separate vulnerabilities that individually do nothing but in sequence produce sophisticated attack outcomes.

Anthropic is not releasing it. They cannot. If this model gets into the wrong hands, it could break the digital infrastructure that the modern economy depends on.

If this is your first time reading, my name is Osiris. I build products, run research, and operate an AI company on my own. I have exited a venture before and now I handle everything end to end. What I share here comes from actual work, shaped by both the constraints and the opportunities of building in the real world.

The Video Nobody Expected

Dario Amodei recorded a video explaining why Anthropic is withholding the model. The language was measured but the implications were not. He said Mythos is as good as a professional human security researcher at identifying bugs. But it is better than any human at chaining vulnerabilities together across long range autonomous tasks.

One security researcher working with Mythos in the preview said he found more bugs in two weeks than he had found in the rest of his career combined. That is not a marginal improvement. That is a category shift.

The benchmarks back it up. On SWE Bench Multimodal, the score went from 27 percent with Claude Opus 4.6 to 59 percent with Mythos Preview. Across every major software coding benchmark, the jump is dramatic. This is not incremental progress. This is a step function.

Project Glass Wing

Anthropic launched Project Glass Wing. A consortium of companies including NVIDIA, AWS, and Azure are working with Anthropic to use Mythos in a purely defensive capacity. The goal is to harden software infrastructure before anyone else develops a comparable capability.

Anthropic also created a 100 million dollar credit fund. Companies can apply for compute credits to use Mythos for finding and patching vulnerabilities in their own systems. The idea is to give defenders a head start.

Thomas Friedman confirmed in the New York Times that representatives of leading tech companies have been in private conversations with the Trump administration about the security implications. This is not a publicity stunt. People who have seen the model are going to the White House and raising alarms.

The Manhattan Project Analogy

The comparison to nuclear weapons keeps coming up. Not because Mythos can cause physical destruction, but because the scale of potential damage to financial and digital infrastructure is comparable to the economic devastation of a nuclear attack.

If a model can find zero day exploits in banking software, in international transfer systems like SWIFT, in cryptocurrency protocols, the financial system becomes vulnerable in ways that no amount of traditional security spending can address. The attack surface is every piece of software that has ever been deployed. And every piece of software has bugs.

The game theory gets dark fast. If Anthropic has this capability, Google probably does too. China has surpassed the United States in AI research papers accepted at top tier conferences and journals. There is a nonzero chance that a comparable capability already exists in Beijing and nobody has announced it.

The proposal from multiple voices on the show is a Manhattan Project for cyber defense. Take the five best cybersecurity minds from each major AI company. Create a 25 person brain trust. Build a system that does its own red teams against critical infrastructure. Nuclear power plants. Banking systems. Military communications. Find the vulnerabilities. Patch them. Do not talk about it.

The Two Tier Economy

There is now a class of companies that have access to Mythos and a class that does not. The companies inside Project Glass Wing can harden their software against threats that the companies outside cannot even detect.

Previously every AI lab was focused on getting their newest model into the world as fast as possible. It was democratic. The best model was available to everyone willing to pay for it. That dynamic just changed.

Polymarket gives a 28 percent chance that Mythos will be publicly available by June 30th. That means three out of four scenarios have the most powerful coding model in the world locked behind closed doors through the summer. In AI terms, that is an eternity.

The question is whether open source catches up before Mythos ships. Meta dropped a new model the same week. It is good. It is not Mythos. The gap between proprietary frontier models and open source has been roughly six months for the past few years. If that gap holds, an open source model with comparable cyber capabilities could exist by the end of the year regardless of what Anthropic decides to do.

Small Language Models As The Counter

While the frontier labs race toward ever larger and more dangerous models, a parallel economy of small language models is emerging. One company offers 39 task specific SLMs packaged under a single API for 8 dollars a month. AT&T rearchitected their entire AI infrastructure to run 90 percent of tasks on small models and cut costs by 90 percent.

The intelligence density of small models keeps increasing. An 8 billion parameter model can do more this year than it could last year because training techniques and architecture improvements filter down from the frontier. By 2030, one estimate suggests 90 percent of common work tasks will be achievable with a 10 billion parameter model running on a laptop.

This creates a paradox. The frontier models become more capable and more dangerous. The small models become capable enough for most tasks at a fraction of the cost. The frontier labs need to justify their pricing against an ever improving floor of cheap alternatives. They may have built the technology that makes their own pricing unsustainable.

The hyperdeflationary pressure is real. Not 10 percent cheaper per year. Possibly 90 percent cheaper per month as compounding improvements in model density, hardware performance, and open source capabilities all stack on top of each other. Compute performance per dollar has improved roughly 40 percent per year across 20 plus AI accelerators released between 2012 and 2025.

The Death Score

A new tool called Death By Claude lets you enter any company URL and receive a defensibility score from 0 to 100. Higher means more replaceable by AI. A textbook company scored 92. A newsletter scored 89. A vibe coding platform scored 78. Peloton scored 32 because you cannot replace a bicycle with code.

The three things that prevent an AI from killing your startup are hardware, network effects, and regulated industries that require human relationships. If your company does not have at least one of those, you are on the clock.

The tool also generates a complete skill file to replace whatever you built. 31 lines of markdown to replicate an entire product. The founder built it because his own startup kept getting mogged by Anthropic releasing competing features. He ran the tool on his own company. It scored 92. So he pivoted.

Every founder should run their company through this exercise. Not as a novelty. As a strategic imperative. If AI can replicate your product in 31 lines, your product is not your moat. Figure out what is before someone else does.

The model finds bugs faster than any human alive. The question is who gets to use it first.