生成が完了しませんでした

もう一度生成を試してください

作成者の再生成待ちです

by EigoBoxAI
作成:2026/04/07 03:04
レベル:上級 (語彙目安:6000〜8000語)
タイプ:リーディング

# 
##

![thumbnail]()

---

When people ask whether “AI that builds AI” is safe, the Claude Code leak offers a sharp answer: only if the boring parts of software engineering are safe too. Claude Code is Anthropic’s agentic coding assistant, designed to read code, edit files, run commands, and automate development tasks from the terminal or IDE. On March 31, 2026, Anthropic accidentally shipped internal Claude Code source in npm package version 2.1.88 through a debugging source map, exposing more than half a million lines of code; the company said no sensitive customer data or credentials were exposed and called it a packaging mistake caused by human error, not a breach. ([docs.anthropic.com](https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/overview))

That irony matters. Anthropic’s security docs describe a cautious design: read-only by default, explicit approval for powerful actions, prompt-injection defenses, and clear warnings that hooks can run shell commands with the user’s full permissions. Yet the failure did not come from some dramatic “rogue AI” scenario. According to Claude Code’s creator, there was “a manual deploy step” that should have been automated better. In other words, the weakness was not futuristic intelligence, but a very old problem: trusting a release process that was faster than its safeguards. ([docs.anthropic.com](https://docs.anthropic.com/en/docs/claude-code/security))

The story also became more dangerous after the leak itself. Reporting indicates this was the second Claude Code source exposure in a little over a year, and within days security watchers were already seeing fake Claude Code repositories and typosquatted packages distributing malware such as Vidar infostealers and GhostSocks proxies. That is the real trap of automated development: one small packaging mistake can ripple outward, first as an intellectual-property leak, then as a supply-chain risk for developers who rush to copy, test, or “unlock” what they find online. So, is AI-made AI safe? It can be—but only if companies treat build pipelines, package contents, and human review as seriously as model safety itself. ([axios.com](https://www.axios.com/2026/03/31/anthropic-leaked-source-code-ai?utm_source=openai))

---