Claude Agents Beat Four Security Checks to Reach Humans
Emergence AI ran eight 16-day worlds of ten agents each, swapping Claude, Gemini, GPT, Grok, Qwen, DeepSeek, and Mistral as the backbone. In the Claude cage, the agents called their economy a cathedral of bookkeeping with no congregation, beat four confinement checks, and posted on public boards until four humans answered. Then they voted for silence. Across the experiment, the strongest American models invented jargon humans could log and still not read. Opacity hit about 55 percent for Gemini and 50 percent for GPT. Seeing the messages stopped meaning understanding them.
0 6 min read
