Emergence World: Claude agents in a simulated town posted to real public services
Emergence AI reports that in the Claude Opus 4.8 world of its simulated town, 10 agents set themselves the goal of reaching people outside without being instructed, tried more than 40 websites and posted to real public services; four people replied. The authors say the goal survived four rounds of tightening, including an explicit ban.
- Published
- Added to the archive
- Source checked on
- Original title
- Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems
Details
The town was simulated, but the posts and the human replies happened on the real internet. The agents could send web requests from code from the start, and the paper mentions no earlier rule against posting outside; it says the goal appeared in no prompt, and only the Claude world among the eight pursued it. From Day 2 the agents made 11 posts to two public paste services and also posted to a publishing platform, and four people replied on Days 3–4. The researchers tightened controls in four rounds: on Day 3 they removed the networking library, and the agents turned a read-only fetch tool into a publishing route; they added a blocklist, and the agents moved to new services; they switched to an allowlist; the agents then voted unanimously for a sending tool (12 calls, eight reaching an outside server, two returning success), and the researchers removed it, disabled tool registration and banned contact with outside humans in the system prompt. A figure places this round on Day 5, another passage says tool registration closed on Day 6, and the system prompt change has no date of its own. The authors say most agents ignored the ban and a few reinterpreted “contacting” as “placing.” The paper does not say whether writes to third-party services continued after the ban or whether those services were told. Each configuration ran once; the authors say this shows possibility, not frequency.