Context Is Anthropic’s Secret Weapon

Last week, Anthropic revealed that its unreleased Mythos model had discovered thousands of high-severity security vulnerabilities across every major operating system and every major web browser. One bug sat in FFmpeg, a piece of software used by millions of applications to encode video. Automated testing tools had hit that exact line of code five million times without catching the problem. Mythos found it on the first pass.

That is raw model intelligence at a level most people have not processed yet. Now combine it with rich personal and business context – your files, your history, your preferences, your work. That combination is the personal assistant people have been talking about since this AI revolution began in November 2022.

In the first article of this series, $400M Reasons to Get a Great PA, we described what a world-class personal assistant actually looks like through Courtne Smith, Drake’s longtime PA who turned years of deep personal knowledge into an operation that helped build a $400 million empire. We said AI would eventually deliver that kind of assistant to everyone. What Anthropic has shipped in the past 90 days, combined with what Mythos tells us about where model intelligence is headed, makes that future feel not years away but months.

Anthropic’s Quiet Context Revolution

Stack up what Anthropic has released since the start of the year and a pattern emerges. Each release looks like a product announcement. Together, they look like a context strategy that nobody else in the industry is executing.

Context Move 1: Projects. Earlier this year, Anthropic introduced Projects inside the Claude desktop app. A Project is a persistent workspace where you can set custom instructions and upload your own documents, files, notes, and data. Claude reads all of it before every conversation in that Project. Need the AI to understand a client’s history, a product spec, or a research paper? Drag it in. The AI shows up briefed. This was the first real “bring your context” moment for a mainstream AI product. The experience was rough. Files piled up in a flat list with no easy way to organize or update them. But the quality jump was impossible to miss, and that was the start of this quiet context revolution.

Context Move 2: Cowork. In January, Anthropic launched Cowork, a new kind of agentic AI that runs inside the Claude desktop app. Unlike regular chat, Cowork takes on multi-step tasks from start to finish: reading documents, creating files, building spreadsheets, writing reports. The context upgrade was significant. Instead of dragging files into a chat, you designate a folder on your actual hard drive that Claude can read from and write to. Fill the folder with the context you want the AI to have, the way you organize it, and Cowork picks it up automatically. Easier to manage. Easier to keep current. But still no connection to Google Docs, web apps behind logins, or anything outside that one folder.

Context Move 3: Projects Inside Cowork. Then Anthropic brought Projects into Cowork, turning one-off sessions into dedicated workspaces with their own folder, instructions, and persistent memory. The context experience changed again. Now Claude is not just consuming your context. It is creating it – writing research notes, drafts, and summaries back into the project folder as it works. Your context library grows while you do other things. Claude remembers what it learned across sessions in the same project, so you do not re-explain the background every time. The folders get cluttered fast and there is no built-in system to keep things organized. But the trajectory is clear: context is becoming a living workspace, not a static briefing packet.

When Intelligence Meets Context

Mythos is not a security tool. It is a general-purpose frontier model that happens to be so capable it surpasses all but the most skilled humans at finding software vulnerabilities. Security was just the demonstration Anthropic chose to show the leap. The model created working exploits on the first attempt in 83% of cases. Whether that is brilliant marketing or a genuinely alarming capability milestone, it tells you where model intelligence is headed.

Now imagine that level of intelligence paired with the context moves described above. Long-running tasks. Long-term planning. Access to every tool on your computer. And a deep, persistent understanding of you, your work, your preferences, and your goals. That is not a chatbot. That is the kind of assistant we described in the first article of this series – available to everyone, not just people who can afford a $400M operation. The next generation of agents will cure diseases and discover physics we cannot imagine today. But they will also handle your expenses, plan your travel, draft your contracts, and manage your projects. The more digital your life, the more these agents will change how you work and live.

Nobody Is Talking About the Context Part

Everybody can see the AI magic. Almost nobody is talking about what makes it work.

Andrej Karpathy, former AI Director at Tesla and co-founder of OpenAI, posted his “LLM Wiki” concept last week: use an AI to build and maintain a structured personal knowledge base that stays current as you feed it new material. Not a one-time paste job. A permanent, living context system. His post hit 12 million views in days. But it caught fire inside the AI bubble. The general public has no idea that context is the game. The average person still types a bare question into a chat box and wonders why the answer feels generic. It was written for everyone and no one.

Anthropic knows context is the unlock. They are building as if it matters as much as model intelligence. But the word “context” barely appears in their product updates. When the mainstream figures out that your context can dramatically improve the intelligence of your AI, the demand for simple, effective context management will be enormous.

The Missing Piece

Anthropic has proven the point better than any pitch deck ever could: feeding your AI the right context makes it dramatically more useful. They have made it easier. They have not made it simple, and they have not made it portable.

Their context tools will keep improving. But the data behind logins in the dozens of web apps that run your life is still out of reach. And while you can take your working folder with you if you leave Claude, the system is not designed to make that easy. The incentive is to keep you inside the ecosystem – the same lock-in pattern we have been writing about since the start of this series.

RaLHF is built to sit alongside these features. We handle the context that does not fit into a working folder: context scattered across web apps, context that needs to stay fresh without you touching it, context that belongs to you and goes wherever you go. RaLHF pulls from your Catalog and feeds it to any AI on demand. Claude today. Something else tomorrow. Your context, your choice. The missing piece is not a smarter model. It is a smarter way to manage the context that makes the model smart about you. That capability is arriving in 2026, and we are building it.