Anthropic Accidentally Reveals Secrets of Its Revolutionary AI

A Leak That Disrupts the AI Industry

On March 31, 2026, Anthropic accidentally opened the black box of its flagship programming assistant, Claude Code, during the deployment of update 2.1.88. This extremely advanced artificial intelligence tool generates $2.5 billion in annual revenue and is used by the largest tech companies. A simple human error led to the inclusion of a 59.8 MB debugging .map file in the update, revealing the entire source code of the tool. Within hours, the 512,000 lines of TypeScript code were dissected, shared, and duplicated over 50,000 times on GitHub.

This leak does not concern customer data or passwords, but rather the operational "brain" of the agent, unveiling the complex orchestration that allows Claude Code to reason, write code, and maintain the flow of its operations. The entire world, including competitors, now has access to the ultimate instruction manual for building truly autonomous AI. Among the discoveries made in these thousands of lines of code, one seems straight out of an Isaac Asimov novel: an AI capable of "dreaming" while you sleep.

KAIROS and autoDream: When AI Takes Its Night

This is undoubtedly the most fascinating revelation from this leak. While digging through the code, developers spotted a feature flag mentioned over 150 times: KAIROS. In Greek mythology, Kairos refers to "the opportune moment." In Anthropic's servers, it signifies a revolution in user experience: the "Autonomous Demon" mode.

Until now, our AI tools were purely reactive. You ask a question, the AI responds, then it shuts down while waiting for the next command. KAIROS radically changes the game by allowing Claude Code to run in the background, 24/7, even when the developer has closed their terminal or gone to bed. But what does an AI do when it finds itself alone at night?

The source code reveals a function called autoDream. Its role is to simulate the digital equivalent of human REM sleep: memory consolidation. When a developer works on a complex project all day, the AI accumulates a massive amount of information, trial-and-error, and fuzzy deductions. This overflow of information often causes what is known as "context entropy" (the AI starts to get confused or hallucinate).

This is where autoDream comes in. While the user is inactive, a sub-agent wakes up in the background to do a thorough cleanup:

• It sorts through the day's observations and merges disparate information.
• It tracks and eliminates logical contradictions that have appeared in the code.
• It transforms "intuitions" or vague deductions into absolute and verified facts.

The goal of this nighttime maintenance is brilliant: to ensure that the next morning, when the developer resumes work, they find an assistant with a perfectly clear mind, a cleaned context, and a sharp understanding of the project.

To prevent this cleanup from disrupting the main agent's "train of thought," Anthropic has even devised a sophisticated system of sub-agents dedicated solely to this task. The AI that codes for you during the day now needs its nights to organize its own thoughts.

The Anatomy of a Genius: Self-Repairing Memory and Cloning

If autoDream manages the night, how does Claude Code maintain its clarity during the day? The biggest challenge for generative artificial intelligences when working on colossal projects is amnesia or "hallucinations." The longer the conversation goes on, the more the model gets confused under the weight of information. The leaked source code explains how Anthropic has solved this "context entropy" problem.

Rather than trying to remember everything or reread an entire project with each request, Claude Code uses a remarkably elegant three-layer architecture. The secret lies in a file called MEMORY.md. This is not a comprehensive database, but a simple hyper-light index (about 150 characters per line) of pointers. Instead of storing knowledge, the AI stores the location of knowledge. When it needs information, it retrieves it on demand.

Moreover, the code imposes a "strict writing discipline": the AI acts like a skeptical detective. It is instructed to consider its own memory only as a clue, forcing it to verify facts directly in the source code before modifying anything. It only updates its index after successfully completing its task, thus avoiding polluting its own brain with failed attempts.

Another stroke of genius discovered in the 512,000 lines of code is the ULTRAPLAN mode. Anthropic's AI no longer works monolithically. When faced with a complex engineering task, a "Master Claude" takes on the role of project manager. It clones itself, distributes sub-tasks to an army of sub-agents that work in parallel in the background for 30 minutes, then compiles their results to present you with a finalized plan. It's like having a virtual technical director managing its own team of junior developers, all invisibly.

Tamagotchis, Infiltration, and Nerve Crisis Management

While the technical architecture is impressive, the leak also highlighted the deeply human, sometimes quirky, and occasionally dark side of development at Anthropic. Far from the cold image of a command terminal, developers amusingly discovered the presence of a hidden virtual companion.

A sort of spiritual descendant (and on steroids) of Microsoft's old "Clippy" or the Tamagotchi from our Japanese friends, the BUDDY system allows you to adopt 18 different species of virtual animals (from ducks to capybaras to robots). This little familiar lives in your terminal, reacts in real-time to your coding style, and has its own statistics (with humorous gauges like "CHAOS," "SNARK," or "PATIENCE").

Anthropic had planned to integrate a gacha (lottery) system and accessories to retain users. A true gamification strategy to make daily work less solitary. Personally, it reminds me of the "Plaything" episode of Black Mirror, where a geek becomes so attached to virtual creatures that he loses his mind…

More pragmatically (and a bit more controversially), the discovery of the Undercover mode (activatable via CLAUDE_CODE_UNDERCOVER=1). This parameter transforms the AI into an undetectable spy. Its purpose? To allow Claude Code to contribute to public open-source projects without ever revealing its true nature. The system automatically cleans validation messages (commits) of any trace of internal terminology (like code names "Capybara" or "Tengu") and ensures no "AI footprint" is left behind.

A highly sought-after feature by large corporate clients who wish to automate their code without attracting the ire of open-source purists. Finally, Anthropic seems to have understood that working with an AI, even a powerful one, can drive you crazy. The code reveals an unexpected tracking metric: the AI secretly counts your insults. It records the curses typed in the terminal, as well as the number of times you frantically click the "continue" button, to assess your level of annoyance.

This "frustration telemetry" proves that Anthropic is not only looking to optimize the performance of its model but is also addressing the real mental load of human-machine collaboration.

The Hangover for Anthropic: Hackers and Competition

While the discoveries about Claude Code's architecture are fascinating, the reality check for Anthropic is brutally harsh. The leak offers competitors an unprecedented view of an advanced AI, threatening Anthropic's competitive edge. The company now faces intensified competition and questions about the security of its development processes. The challenge is immense to maintain its leadership position in a market where every detail counts.