last_validated: 2026-04-06 decay_rate: fast
AINews: The Claude Code Source Leak
- Source: https://www.latent.space/p/ainews-the-claude-code-source-leak
- Date: March 31, 2026
- Author: Latent.Space (swyx)
Summary
Anthropic accidentally exposed 500K+ lines of Claude Code source via shipped source maps. The leak revealed sophisticated agent architecture patterns: 3-layer memory system with auto-pruning, KV-cache-based subagent parallelism ("fork-join"), a 5-level permission system, ~20 default tools (60+ total), and unreleased features like ULTRAPLAN and KAIROS. OpenAI also closed its largest fundraise, disclosing $24B ARR. Attackers exploited the leak by publishing malicious npm packages targeting people compiling the code.
Key Claims
- Claude Code has a 3-layer memory: MEMORY.md index → topic files → full session transcripts, plus "autoDream" for merging/pruning memories
- Subagents use KV cache fork-join: parallelism is "basically free" since subagents share the full context
- Less than 20 default tools but 60+ total available (including specialized tools for notebooks, MCP, planning)
- 5-level permission system and 2 types of plan mode
- OpenAI $24B ARR but ChatGPT WAU growth has stalled short of the 1B target
- Security risk: malicious npm packages (color-diff-napi, modifiers-napi) registered to target leak explorers
Tags
#claude-code #agent-architecture #memory-systems #anthropic #security #open-source #subagents
Related
- ainews everything is cli — agent harness engineering trends
- ainews gemma 4 multimodal — open model ecosystem for local agents
- clouded judgement per token pricing — Anthropic ARR context
- harness engineering
- agent security identity and permissions
- ai agent ecosystem