The Mythos Moment
On April 7, 2026, Anthropic announced Claude Mythos Preview. The headline was striking on its own: a single model, running autonomously, identified thousands of previously unknown security vulnerabilities across every major operating system and web browser. One of them, a remote code execution flaw in FreeBSD's NFS implementation, had sat in the codebase for 17 years. It survived peer review, manual audits, and millions of automated tests. Mythos found it in hours, then wrote a working exploit.
A few weeks later, security researchers reported that Mythos had also surfaced a 27-year-old TCP SACK vulnerability in OpenBSD and a 16-year-old flaw in FFmpeg. Anthropic chose not to release the model commercially. Instead, the company launched Project Glasswing, a consortium of major tech firms working through responsible disclosure of Mythos's findings.
The reaction split predictably. Some treated it as a generational leap, the moment AI moved from "helpful assistant" to "autonomous agent with offensive capabilities." Others, including The Ringer, asked whether Mythos could in principle destroy the internet. Both reactions miss the more useful question for anyone who's not running a security team.
The question is this: if a model this capable still has no idea who you are, what does that tell you about where the durable value lies?
What Mythos Can Do, and What It Can't
It's worth being concrete about what Mythos actually represents, because the gap between hype and reality is where useful thinking lives.
According to Anthropic's red team report and InfoQ's coverage, Mythos Preview shows the largest gains over Claude Opus 4.6 in mathematics, long-context reasoning, software engineering, and cybersecurity. On standard benchmarks it sets new records. On adversarial security tasks it operates with a degree of autonomy that previous models couldn't sustain. Give it a target, and it can plan, probe, and iterate without further instruction.
But notice what's missing from that list.
Mythos doesn't know which book you read on a flight in 2021 that quietly reshaped how you think about your career. It doesn't know that a single paragraph from a Tyler Cowen interview made you reconsider an investment decision. It doesn't know the three highlights from a Naval Ravikant tweetstorm that you've returned to four times this year. It doesn't know your reading speed, your tolerance for ambiguity, your taste in counterargument, or the specific way you tend to misread certain kinds of evidence.
Here's the asymmetry. Mythos was trained on the public internet plus whatever proprietary data Anthropic could license. So was every other frontier model. Public information is the input every lab has access to. It's table stakes. What no model has, by default, is the trace of your specific mind moving through information over time.
That trace lives in three places. It lives in your head, where it's lossy and prone to forgetting. It lives in scattered tools (notes apps, screenshot folders, browser bookmarks) where it's effectively invisible to AI. Or it lives in a structured form that AI can actually consume.
The third option is what changes everything.
Why Personal Context Is the New Moat
In strategy, a moat is a structural advantage that competitors can't easily replicate. Coca-Cola's moat is brand. Google's moat is the link graph plus user click data. AWS's moat is switching costs.
For individuals working with AI, what's the analogue?
It isn't access to better models. Frontier capability is converging fast and the gap between the best closed model and the best open one is now measured in months. It isn't prompt-writing skill, because that's a teachable craft and the leverage is finite. It isn't even raw domain expertise, since AI can compress the gap between novice and expert on most explicit-knowledge tasks.
The moat is the part of your cognition that's legible to AI and unavailable to anyone else. Call it personal context.
This was already true before Mythos. What Mythos clarifies is that no amount of model capability erases the gap. A more powerful model with no context about you doesn't produce better personalized output. It produces more confident generic output, which is worse, because confidence raises the cost of catching errors.
Think about what happens when you ask a generic AI for advice on a career decision. It pattern-matches on millions of similar prompts and gives you a response shaped like the average. If you've never told the model that you optimize for optionality, that you've already been through one founder burnout, that the specific industry you're in is shifting in ways the training data doesn't reflect, it can't possibly weigh those factors. It will give you fluent, plausible advice that's calibrated to no one.
Now imagine you've spent three years feeding it your highlights from books and articles, your notes from podcasts, your annotations on your own writing. Suddenly the same model gives you advice that references your stated priorities, the framework you highlighted from a Patrick O'Shaughnessy interview, the specific failure mode you flagged when reading about another founder's burnout. Same model. Different output. The delta is context.
This is the same logic that drove building a second brain and the Tiago Forte framework, pulled into sharper focus by the Mythos release. The point of capturing isn't retrieval by you. It's feeding a system that becomes more valuable as both your context and the underlying model get better.
Highlights as the Atomic Unit of Context
Not all context is equally useful to AI. The format matters more than people realize.
Consider the difference between three ways of capturing the same insight from a book.
You could read the book and remember roughly that "the author argued something about feedback loops." That's useless to AI because it can't be retrieved as a discrete fact. It's also lossy, because you'll forget the nuance within a week.
You could copy the entire chapter into a notes app. That's worse than useless. AI now has to scan thousands of irrelevant words to find the sentence that mattered to you. Worse, you've stripped the source attribution, so the model can't verify or extend the claim.
Or you could highlight the specific sentence that struck you, leave it tied to the source URL or book chapter, and optionally add a one-line note about why it mattered. That highlight is now a context atom. It's pre-filtered (you chose it), short enough to load cheaply into a prompt, source-attributed for verification, and small enough that thousands of them fit in a single retrieval call.
This is why Glasp's web highlighter is structured the way it is. A highlight is the smallest unit of attention you can capture while still preserving everything AI needs. Multiply that by a few hundred or a few thousand across the books, articles, and PDFs that actually shaped your thinking, and you have a personal corpus that no labs's training data covers.
Add Kindle highlights on top, which often represent your deepest, most considered reading, and the corpus deepens.
Add YouTube Summary timestamps from videos you found dense enough to actually study, and you've now captured a layer of context that almost no one bothers to preserve (video knowledge is famously hard to extract from after the fact).
You've turned what most people lose into something that compounds.
| Capture Method | Searchable by AI | Source-attributed | Pre-filtered for relevance | Cost to AI |
|---|---|---|---|---|
| Memory only | No | No | Yes, lossy | N/A |
| Screenshots | Poor (OCR varies) | No | Yes | High |
| Full-text bookmarks | Yes | Yes | No | High |
| Hand-written notes | No | Sometimes | Yes | N/A |
| Highlights with source | Yes | Yes | Yes | Low |
The format isn't a nice-to-have. It's the difference between a context system AI can use and a pile AI has to ignore.
The Personal Context Stack
Once you start thinking in terms of personal context as a moat, it helps to see it as a stack, not a single thing.
Layer 1: Raw highlights. The sentences and clips you marked. These are the smallest, cheapest atoms. They answer the question "what did this person find worth slowing down for?"
Layer 2: Annotations. A short note attached to a highlight that captures the why. Not "this is interesting" but "this contradicts X" or "useful for the Q3 decision." Annotations add signal that the highlight alone can't carry.
Layer 3: Profile and stated goals. Who you are in the world. What you're trying to learn, build, or decide. This is the equivalent of a system prompt for a personal AI. It tells the model how to interpret everything in layers 1 and 2.
Layer 4: Social context. What others in your network found worth highlighting in the same sources. This is where Glasp's community feed becomes useful. Seeing five people you trust highlight different sentences from the same essay tells you something about the essay that no single person's highlights would.
Layer 5: AI chat over the stack. This is where it all becomes operational. Tools like Glasp's AI chat let you query your stack directly. "What did I highlight about decision-making under uncertainty?" "Pull every highlight from my last six months that touched on hiring." "Find the contradiction between what Marc Andreessen said in this interview and what I highlighted from Peter Thiel two years ago." Without the stack, those questions don't even make sense.
What's interesting is that each layer compounds the others. A profile without highlights is a generic prompt. Highlights without annotations are atoms without bonds. Social context without your own corpus is just a feed. AI chat without a stack to query is the same generic chatbot everyone else has.
The full stack is the only thing on this list that's genuinely yours.
What Glasswing Teaches Individuals
Project Glasswing, Anthropic's response to Mythos's offensive capability, is worth studying as a strategic pattern. It also happens to be a near-perfect illustration of the personal context principle applied at industry scale.
Anthropic could have released Mythos publicly. They chose not to. They could have published the discovered vulnerabilities openly. They chose not to. Instead, they built a curated network of trusted parties (Microsoft, Apple, Google, and others) and shared findings inside that network, with a coordinated disclosure timeline.
The reasoning, as covered in CETaS's analysis from the Alan Turing Institute, is straightforward. Public release of capability without context produces chaos. Curated release of the same capability inside a network that has the context to act on it responsibly produces fixed systems.
Apply that to individuals. A frontier AI without your personal context is the offensive Mythos, capable but unaligned with your specific situation. The same AI with access to a curated personal stack is the Glasswing version, capable and contextual. The model doesn't change. The surrounding network of context does.
This is the same principle behind collective intelligence. The value isn't in any single piece. It's in the curation and the relationships between pieces.
For an individual learner, the practical translation is: stop trying to keep up with every AI release. Start building the context layer that makes any release useful to you.
Building Your Moat Without a Lab
You don't need a research team or a frontier model lab to build personal context. You need a system that's low enough friction that you'll actually use it, and structured enough that AI can read it later.
A few principles that hold up regardless of which tool you use.
Capture at the source, not later. The window where a sentence still feels important is short. If you bookmark it now and plan to take notes "later," you'll lose 80% of the signal. Highlighting in-browser, on the page, at the moment of attention, is the only way to preserve why it mattered. Glasp's web highlighter was built around this constraint.
Annotate sparingly but specifically. A blank highlight is useful. A highlight with "this is great" is not, because it doesn't explain why. A highlight with "contradicts the framework I used last quarter" is gold. One annotation per ten highlights is fine. Make those annotations count.
Preserve sources religiously. A quote without a source is unverifiable. Future-you will encounter the highlight and not remember whether the author was credible. AI will treat it with the same suspicion. Source attribution is a small habit with outsized leverage.
Treat video and audio as first-class. Most personal context systems silently exclude video, which is a problem because half of serious learning now happens through podcasts and YouTube. YouTube Summary with timestamp-anchored highlights solves this. The video equivalent of a book highlight is a 30-second clip with a quote pulled from the transcript.
Revisit on a slow cadence. Highlights you never re-read are still useful (AI can find them), but highlights you revisit gain a second layer of signal: the fact that you came back. This is the synthesis loop (capture, revisit, connect) that turns raw context into something more like reasoning.
Trust your network for what you can't capture yourself. No one reads everything. The community feed is where you see what people you respect found worth marking. Treat it as a parallel pipeline that fills in the corners of your own reading.
None of this is heroic. It's a small set of habits that, applied steadily, produce a context layer that no model will have unless you build it.
The Capability Gap Will Widen
There's a tempting line of thinking that says: AI will get good enough that personal context won't matter. Just describe what you need, and a smart enough model will figure out the rest.
The Mythos release is evidence against that view, not for it.
Here's why. Capability and context aren't substitutes. They're complements. A smarter model with no context about you is a more powerful tool aimed at the wrong target. Imagine handing Mythos's full capability to someone who doesn't know your business, your priorities, or your past mistakes. The output would be confident, sophisticated, and wrong in ways that are hard to detect because the surface fluency is so high.
This is the AI thinking trap at its most dangerous level. When the model is bad, you notice. When the model is great but generic, you don't, and you act on its output anyway.
The way out isn't to slow down the models. They're getting faster regardless. The way out is to build the personal context layer that lets you actually use them. The labs are pouring resources into capability. Almost no one is pouring resources into your context. That asymmetry is your opportunity.
In two years, the people who have a working personal context stack will have AI that operates as a coherent extension of their thinking. The people who don't will have AI that operates as a confident stranger. Both groups will be using the same models.
Frequently Asked Questions
Is Claude Mythos available for public use?
No. As of May 2026, Anthropic has not released Mythos Preview commercially. The model is being used internally and through Project Glasswing for responsible vulnerability disclosure. Anthropic has cited cybersecurity risks as the reason for withholding general access. This is part of why the model is interesting as a case study: it represents capability that exists but isn't widely accessible, which makes the "what would you do with it if you had it" question worth taking seriously.
Doesn't this make personal context dependent on a single AI provider?
No, if you structure your capture correctly. Highlights, annotations, and notes are portable. They're just text with source attribution. If you store them in a system that lets you export, you can move them between AI providers as the frontier shifts. The thing to avoid is any tool that locks your context inside a proprietary format you can't extract. Glasp lets you export your highlights, which is the property you want regardless of which AI you end up using.
How is this different from just using ChatGPT's memory feature?
ChatGPT's memory is an opaque, provider-controlled summary of what it has inferred about you from conversations. You can't audit it, can't easily edit it, and can't port it to another model. Personal context, in the sense this article uses it, is structured data you own (highlights, notes, profiles). You can inspect it, edit it, version it, and feed it to any AI. One is a feature of a product. The other is a moat that survives product changes.
What if I haven't been highlighting for years? Is it too late?
It's never too late, but you can't compress time. The system you start feeding today will be useful in months and indispensable in years. There's a Charlie Munger line about the best time to plant a tree being twenty years ago and the second best time being now. The same applies. People who started Glasp three years ago have a meaningful context advantage today. People who start this month will have that same advantage in 2029.
Won't future AI just read my entire browser history and infer all this automatically?
Some of it, maybe. But two things matter. First, raw browser history is mostly noise. The signal-to-noise ratio is much lower than highlights, which are pre-filtered by the most expensive resource in your life: your attention. Second, even if a future AI could reconstruct your context from history, the provider holding that data isn't necessarily one you want to depend on. Owning your context layer is a hedge against the future shape of the AI industry, not just an efficiency gain.
How does YouTube Summary fit into this?
Video and audio are usually invisible to personal context systems, which is a major gap because so much serious learning happens through them now. YouTube Summary extracts a transcript, surfaces key moments, and lets you highlight specific timestamps. The output is functionally identical to a book highlight: a short quote, a source, a moment in time. That means your YouTube learning becomes first-class context, queryable by AI alongside everything else you've read.
Doesn't more context mean more AI prompt cost?
Yes, but the cost has been collapsing fast, and context windows have been expanding. Two years ago, loading 1,000 highlights into a prompt was expensive. Today it's trivial. By the time you've built a corpus large enough to matter, the cost of using it will be lower than the cost of not having it. The bigger risk is not having the context, not having too much of it.
Conclusion: Start Feeding the System That Will Know You
Claude Mythos is the clearest illustration yet of what frontier AI can do. Find a 27-year-old bug in code that survived decades of human review. Generate working exploits with no human prompting. Do all of it in hours. The capability is real, it's accelerating, and it's not going to slow down.
What that capability doesn't include, and won't include by default, is any model of you. Your priorities, your reading history, the specific sentences that shaped how you think. That part of the picture is invisible to even the most capable model unless you make it visible.
The work is small but it doesn't compress. One highlight at a time. One annotation when the context warrants it. One short profile that tells AI how to weigh everything else. One pass through your network's highlights when you want a perspective you didn't generate yourself. None of it is heroic. All of it compounds.
The Glasswing pattern is the right mental model for individuals too. Mythos plus a curated context network produces useful, scoped, responsible output. Mythos with no context is a capability looking for the wrong problem. Apply the same logic to yourself. The frontier will keep moving. What stays yours is the corpus you've built around it.
Start with one highlight today. The system that knows you in 2028 is the one you start feeding now. Glasp's web highlighter, YouTube Summary, and AI chat over your highlights are designed around exactly this premise: capture cheaply, attribute everything, and let AI do the heavy retrieval on a corpus that no one else has.
The model isn't your moat. Your context is.