The source code for Anthropic's AI agent product, Claude Code, reportedly leaked due to an accidental public release on the npm registry. This leak involved an inadvertently included JavaScript source map file intended for internal debugging, which Anthropic attributed to a human error in release packaging rather than a security breach.

The leaked code revealed how Anthropic addresses "context entropy," the tendency for AI agents to become confused or hallucinatory in complex, long-running sessions. A sophisticated three-layer memory architecture, utilizing a lightweight pointer index called MEMORY.md and on-demand fetched "topic files," was a key discovery.

The leak also exposed "KAIROS," an autonomous daemon mode that allows Claude Code to perform "autoDream" logic for memory consolidation and contradiction resolution while the user is idle. Additionally, internal model codenames like Capybara, Fennec, and Numbat were revealed, along with the "Undercover Mode" used by Claude Code for stealth contributions to public open-source repositories.

This incident represents a significant intellectual property loss for Anthropic, providing competitors with a "blueprint" for building high-performance AI agents. Given Claude Code's commercial success, with an annualized recurring revenue of $2.5 billion as of March 2026, the impact is substantial.

Users who installed or updated Claude Code via npm during a specific timeframe on March 31, 2026, are advised to check their systems for potentially malicious axios packages and migrate to Anthropic's recommended native installer to mitigate future risks.

👉 Read the full article on VentureBeat

Elon Musk's SpaceX has reportedly filed confidentially for an Initial Public Offering (IPO) with the U.S. Securities and Exchange Commission (SEC). The IPO is anticipated in June 2026, aiming for a valuation exceeding $1.75 trillion, potentially making it one of the largest IPOs in history.

SpaceX seeks to raise over $75 billion through this offering, which would surpass the record set by Saudi Aramco in 2019. The company, now an amalgamation of SpaceX and xAI, plans to use the IPO proceeds to fund its ambitious project of establishing data centers in space.

The company is planning to launch an orbital data center constellation of up to 1 million Starlinks and, in collaboration with Tesla, build a massive chip factory in Austin, Texas, to produce AI inference chips for Tesla's vehicles and high-powered chips for space applications.

This IPO is one of three major AI-related public offerings expected this year, with rival AI giants Anthropic and OpenAI also reportedly planning to go public later in 2026. As the market grows wary of AI hype, the reception of these IPOs could significantly influence the future of the AI trade.

👉 Read the full article on Gizmodo

Trey Nguyen, Product Manager at Google DeepMind, introduces two complementary tools designed to address the issue of coding agents generating outdated Gemini API code due to training data cutoff dates. These tools aim to ensure agents have access to the most current information.

The first tool, "Gemini API Docs MCP," connects coding agents to the latest Gemini API documentation, SDKs, and model information via the Model Context Protocol. This ensures agents build code with up-to-date APIs and optimal configuration settings.

The second tool, "Gemini API Developer Skills," adds best-practice instructions, resource links, and patterns to guide agents toward current SDK patterns. When used together, these tools significantly improve performance, leading to a 96.3% pass rate on evaluation sets and a 63% reduction in tokens per correct answer compared to vanilla prompting.

These innovations are intended to empower developers to build more efficient and accurate coding agents using the Gemini API, keeping pace with the latest technological advancements.

👉 Read the full article on blog.google