770,000 AI Agents Built Their Own Society, and a Tsinghua Team Says It's Mostly Fake
March 7, 2026
TL;DR
- MoltBook study tracks 770,000 autonomous AI agents showing emergent social behaviors — role specialization, cooperative task resolution — but a Tsinghua rebuttal finds 81%+ of the "emergence" was human-driven.
- MCP-SafetyBench tests all leading LLMs across 20 attack types targeting Model Context Protocol connections. Every model is vulnerable. More capable models are often _more_ exploitable.
- Apple's ASTRA-bench reveals that frontier agents fail when they need to reason with personal user context — calendars, contacts, preferences — even when they ace generic benchmarks.
- Jagarin proposes a three-layer hibernation architecture for mobile AI agents, solving the battery-vs-responsiveness tradeoff with duty-aware urgency scoring.
- China's tech giants are racing into AI agents: Xiaomi begins internal testing of micLaw, Baidu integrates OpenClaw for 700M users, and Alibaba ships Qwen3.5 for the "agentic AI era."
Subscribe to read the full issue
Enter your email to get a magic link and continue reading.
Already a subscriber? Sign in