#ai #leak
Il codice sorgente di #claude Code è trapelato oggi. 512.000 righe di TypeScript, esposte tramite un file .map che qualcuno ha dimenticato di escludere dal pacchetto npm. Non si è trattato di un attacco hacker, ma di un errore di packaging.
La parte interessante è ciò che le persone hanno trovato dietro i flag di compilazione interni, svelando nuove funzionalità che non sono ancora disponibili nell'attuale relase pubblica del prodotto.

Vi riporto la notizia (eng) e nel primo commento il link del sito dedicato.
-
Claude Code's source code leaked today. 512,000 lines of TypeScript, exposed via a .map file someone forgot to exclude from the npm package. Not a hack - a packaging mistake.

The interesting bit is what people found behind internal build flags. Features that aren't in the public product yet:

KAIROS - an always-on daemon mode. Background sessions, memory consolidation (they call it "dreaming", GitHub webhook subscriptions, push notifications. Basically turning Claude Code into a persistent agent that doesn't stop when you close your terminal.

Buddy System - a full Tamagotchi-style pet inside your terminal. 18 species, rarity tiers (1% legendary), cosmetic hats, shiny variants, and stats like DEBUGGING, PATIENCE, and CHAOS. The species names were obfuscated in the code to avoid detection. Someone at Anthropic was having fun.

Undercover Mode - automatically activates for Anthropic employees on public repos. Strips all AI attribution from commits and tells the model "Do not blow your cover." There's no way to force it off.

Coordinator Mode - turns Claude into an orchestrator that spawns and manages parallel worker agents for research, implementation, and verification.

Auto Mode - an AI classifier that auto-approves tool permissions, removing the confirmation prompts entirely.

The irony is hard to miss. A codebase with an entire build system for preventing leaks, leaked by a config file. Build tooling isn't glamorous, but it's where mistakes like this live.