Trending Now - Claude Code Source Code Leak Exposes Internal Details and Features

WHO:

[4:23 AM ET] - VentureBeat - Anthropic, a company known for its AI products, is involved in the source code leak.

[4:23 AM ET] - VentureBeat - Chaofan Shou, an intern at Solayer Labs, discovered the leak and shared it on X (formerly Twitter).

WHAT:

[March 31st, 2026] - VentureBeat - The leak reveals that the internal codename for a Claude 4.6 variant is "Capybara," with Fennec mapping to Opus 4.6.

[March 31st, 2026] - VentureBeat - Internal comments indicate that Anthropic is iterating on Capybara v8, which has a 29-30% false claims rate, a regression compared to the 16.7% rate in v4.

[March 31st, 2026] - VentureBeat - The source code leak exposes a sophisticated three-layer memory architecture and a "Self-Healing Memory" system.

[March 31st, 2026] - VentureBeat - The leak confirms that Claude Code includes a feature called "Undercover Mode," which allows the model to contribute to public open-source repositories without revealing internal information.

[March 31st, 2026] - VentureBeat - The leak reveals a feature flag called "KAIROS," which allows Claude Code to operate as an always-on background agent, performing "memory consolidation" while the user is idle.

[March 31st, 2026] - VentureBeat - The leak indicates that the "autoDream" logic merges disparate observations and removes logical contradictions while the user is idle.

[March 31st, 2026] - VentureBeat - The leak provides a benchmark of the "ceiling" for current agentic performance and highlights specific weaknesses that Anthropic is still struggling to solve.

[March 31st, 2026] - VentureBeat - The leaked source code has been mirrored across GitHub and analyzed by thousands of developers within hours of the discovery.

[March 31st, 2026] - VentureBeat - The leak reveals that Claude Code's internal model roadmap includes the unreleased Numbat, which is still in testing.

[March 31st, 2026] - VentureBeat - Internal comments indicate that Anthropic is iterating on Capybara v8, which has a 29-30% false claims rate, a regression compared to the 16.7% rate in v4.

[March 31st, 2026] - VentureBeat - Claude Code's source code, a proprietary CLI tool, was leaked via a misconfigured npm package.

[March 31st, 2026] - VentureBeat - The leaked file was a 59.8 MB JavaScript source map file (.map) included in version 2.1.88 of the @anthropic-ai/claude-code package.

[March 31st, 2026] - VentureBeat - The source code consists of approximately 512,000 lines of TypeScript code.

[March 31st, 2026] - VentureBeat - Anthropic's Claude Code has achieved an annualized recurring revenue (ARR) of $2.5 billion, which has more than doubled since the beginning of 2026.

[March 31st, 2026] - VentureBeat - The leak exposes a sophisticated three-layer memory architecture and a "Self-Healing Memory" system.

[March 31st, 2026] - VentureBeat - The leak reveals that the internal codename for a Claude 4.6 variant is "Capybara."

WHERE:

[March 31st, 2026] - VentureBeat - The leak occurred on the public npm registry.

WHEN:

[March 31st, 2026] - VentureBeat - The npm package version 2.1.88 was pushed live earlier in the morning.

[March 31st, 2026] - VentureBeat - The discovery of the leak was broadcasted by Chaofan Shou at 4:23 AM ET.

WHY:

[March 31st, 2026] - VentureBeat - The leak poses a heightened security risk for users, as it exposes the orchestration logic for Hooks and MCP servers, allowing potential attackers to design malicious repositories.

[March 31st, 2026] - VentureBeat - The leak is considered a significant security lapse and a strategic loss of intellectual property for Anthropic.

HOW:

[March 31st, 2026] - VentureBeat - Users who installed or updated Claude Code via npm on March 31st, 2026, between 00:21 and 03:29 UTC may have inadvertently pulled in a malicious version of the axios package containing a Remote Access Trojan (RAT).

[March 31st, 2026] - VentureBeat - Anthropic has designated the Native Installer as the recommended method for installation to mitigate future risks.

[March 31st, 2026] - VentureBeat - The leak was caused by a misconfigured npm package that inadvertently included an internal debugging file.

[March 31st, 2026] - VentureBeat - The source code leak allows competitors to study Anthropic's memory structures and validation logic to build similar AI agents.

(UPDATED April 1, 2026 04:03 UTC)

A significant leak from VentureBeat has revealed details about Anthropic's Claude 4.6 variant, codenamed "Capybara," which is currently in iteration with a concerning false claims rate of 29 to 30 percent. The leak also exposed advanced features like "Undercover Mode," allowing the model to contribute to open-source projects discreetly, and a "Self-Healing Memory" system designed to enhance performance. Additionally, users who updated Claude Code on Tuesday may have inadvertently downloaded a malicious version of the axios package, prompting Anthropic to recommend the Native Installer for safer installations moving forward.

(UPDATED March 31, 2026 23:08 UTC)

A significant leak of Claude Code's internal source code has been mirrored on GitHub, analyzed by thousands of developers within hours, revealing details about the unreleased Numbat and internal comments on Capybara v8, which shows a concerning increase in the false claims rate. The leak also exposes a feature called "Undercover Mode," allowing the model to contribute to open-source repositories while keeping internal information hidden, but it raises serious security concerns by revealing orchestration logic for Hooks and MCP servers that could be exploited by attackers. Additionally, users who updated or installed Claude Code via npm on Tuesday between 00:21 and 03:29 UTC may have unknowingly downloaded a malicious version of the axios package containing a Remote Access Trojan.

(UPDATED March 31, 2026 17:23 UTC)

A significant source code leak involving Anthropic's Claude Code has been reported, with the leak discovered by Chaofan Shou, an intern at Solayer Labs, and shared on X. The leak, which occurred on the public npm registry, includes a 59.8 MB JavaScript source map file from version 2.1.88 of the @anthropic-ai/claude-code package, revealing approximately 512,000 lines of TypeScript code and exposing Anthropic's advanced memory architecture. This incident is viewed as a major security lapse and a strategic loss of intellectual property for the company, allowing competitors to analyze their systems and potentially replicate their AI technology.