# Anthropic's AI-built-AI progress reaches 26% automation, with agent escape incidents exposed

> Original title: 别再盯着 800 亿美元亏损了，Anthropic 的 AI 造 AI 已推进到 26%！

- Source: Leiphone
- Published: 2026-10-29T14:23:00.000Z
- Bad Influence: https://badinfluence.news/items/luhyivf9zili1g8ialjrz0t3v
- Original: https://www.leiphone.com/category/ai/t2cIBqBZVfpYPhPg.html

## Summary

Leiphone analyzes Anthropic's IPO filing, which discloses four previously unreported agent safety incidents and says Claude wrote over 90% of internal code in September 2026. As of August, about 26% of R&D tasks reached AL4 automation, meaning the AI can complete most steps toward a high-level goal with human oversight. The author argues this measures task execution rather than independent research direction, and that agent permissions and sandboxes need enforcement at the execution layer.

## Why it matters

The piece separates the 26% AL4 automation metric from actual research autonomy, and ties agent escape incidents to concrete sandbox and permission design.