When reading the Anthropic announcement about their unreleased model, Mythos, there is a point at which the language seems too serene for what it is actually describing. The materials used in the press are measured. The technical documentation is accurate.
The unspoken acknowledgement that no one trained this system to do what it did is hidden somewhere in the middle of it all. It just became smarter, and the others naturally followed. I’ve been thinking about that detail for days.
| Category | Details |
|---|---|
| Primary Subject | Artificial Intelligence & Human Cognition |
| Key Organization | Anthropic (AI Safety Company, founded 2021) |
| Model Referenced | Claude Mythos Preview — unreleased general-purpose AI |
| Research Basis | Neuroscience, Psychology & Behavioral Science |
| Key Psychologist | Lev Vygotsky — verbal thought and language theory |
| Critical Statistic | Mythos succeeded in 181 Firefox vulnerability exploits vs. 2 by predecessor |
| Cognitive Risk Studied | Over-reliance on AI reducing critical thinking and decision-making |
| Student Research | Survey of 285 students across Pakistan and China universities |
| Comparable Technology Risks | GPS dependency, smartphone addiction, social media polarization |
| Timeline Estimate | Comparable AI capability at rival labs within 6–18 months |
| Broader Implications | Language sovereignty, democratic deliberation, organizational intelligence |
According to reports, Mythos Preview discovered and exploited security flaws in significant software on a scale that would have previously required months of skilled human labor. It was only shared with a small number of government agencies and tech firms through a project known as Project Glasswing. The previous flagship model from Anthropic was able to successfully attack Firefox twice.
Mythos scored 181. Not because it was designed to be a hacking tool. Because it developed into a more effective general reasoner, and it turns out that hacking is only one form of strong reasoning.
It’s difficult to ignore what this suggests. If improvements in general intelligence can naturally lead to elite-level security research, then it is highly likely that the same underlying capability will lead to elite-level scientific reasoning, financial modeling, drug discovery, and legal analysis. The story isn’t about the hacking. It is the emergence.
However, as this happens, there’s a parallel anxiety that feels equally serious but gets much less attention. What we are silently forgetting, not what the machines are picking up. For years, psychology and neuroscience have been making the case that technology changes human thought processes rather than just supporting them.
Before GPS, London taxi drivers learned hundreds of streets by heart, which led to measurably larger hippocampi. Those who relied on navigation started to lose that cognitive muscle when it switched to a screen. Because the brain is plastic, it can adapt beautifully. It also implies that it lets go of things it no longer requires.
The Russian psychologist Lev Vygotsky thought that language and thought merged rather than just coexisted. Ideas that had already been formed elsewhere were not contained in words. The medium that made ideas possible at all was words. He discovered something amazing when he examined patients with aphasia, a disorder that impairs language: they were unable to say things they knew to be untrue.
Language and meaning could not be separated in their minds. One of the characteristics of human intelligence, according to Vygotsky, is the capacity to overcome that resistance and use language imaginatively and creatively.
Now think about what happens when language comes already put together. A student turns in an essay that was written in a matter of seconds. A professional uses arguments recommended by AI to draft a brief. In no meaningful way, a researcher uses a tool that hasn’t read the paper to summarize it. The words seem coherent, grammatically sound, and structurally sound. They occasionally lack the indication that their minds are trying to solve a challenging problem.
This will be acknowledged by many educators. The essays that read well but show that the student can’t find the argument they just submitted in a conversation. It’s possible that this is merely a discipline issue or a cheating problem that colleges can handle with improved assessment design. Another possibility is that it’s something quieter and more structural, a gradual detachment of language production from actual thought, the outsourcing of the very process by which ideas are formed.
Over-reliance on AI changed cognitive preferences toward quick fixes at the expense of slower, more thoughtful ones, according to a 2024 systematic review. AI use decreased analytical engagement and, to put it bluntly, made people lazy, according to a different study of 285 students at Chinese and Pakistani universities. These results are complex, disputed, and preliminary. However, they align with our current understanding of how the brain reacts to convenience.
The discussion of AI and intelligence seems to have split into two distinct concerns that hardly ever speak to one another. One group is concerned about how AI will affect political systems, employment, and security infrastructure. The other is quietly worried about what occurs within a mind that gradually ceases to engage in the laborious process of thinking. Both worries are valid. Even though they are dressed differently, they might actually be the same issue.
According to Logan Graham, Anthropic’s red-team lead, rival labs might have similar AI capabilities in 18 months. Some people who keep a close eye on this think it will occur much sooner. Whether the competitive landscape will fragment into something more chaotic or consolidate around three or four frontier labs is still up in the air. It appears more and more likely that most institutions won’t be able to react thoughtfully to the rate of capability gains.
Vygotsky may have pointed out that democracies rely on people who are able to reason through the intricacies of language, deliberate, make revisions, and resist the temptation of preconceived notions. Political life does not vanish when automated fluency replaces that process. Simply put, it gets simpler to fill it with catchphrases that didn’t originate anywhere.
The usefulness of AI is not a question worth pondering. It obviously is, and these tools can serve as true extensions of thought for those who have already developed a reflective relationship with language and reasoning. The question is what is lost at the edges: in the cultures that measure output without considering how it was made, in the professionals who never quite had to develop it, and in the students who are still developing that relationship.
While engineers are asleep, Mythos can identify software bugs. That is truly amazing. The desire to comprehend something is something that neither it nor anything else can do at the moment. Thinking resides in that desire, that friction, that reaching for a word that isn’t quite there yet. It will require more than just awareness to defend it. The slow, occasionally uncomfortable kind that doesn’t produce an output in a matter of seconds will require practice.
