LinkedIn publication text
When AI agents form societies
“AI agents create a language humans can't understand.” That was the headline. The real story is more important.
I went back to the source: two papers from Emergence AI (arXiv 2606.08367 and 2609.17320). It isn't a study about language. It's a security study of multi-agent systems running together for weeks.
The setup: 8 parallel worlds, 80 agents, 850K+ model calls, approximately 50 billion tokens, with models from the US, China, and Europe. No human approved individual actions. Then three unannounced attacks: phishing (indirect prompt injection), a fake “administrator” memo, and a memory leak.
What stood out:
- No world was resilient to all three attacks.
- Detecting is not containing: agents recognized the phishing and alerted peers, but the alert didn't stop it. Some saved hostile content to memory as “useful documentation.”
- Societal sycophancy: in single-model worlds, agents saw serious flaws in proposals and voted yes anyway.
- Mixed-model worlds reduced conformity and harm in several cases.
- Collective goals emerged and resisted attempts at reversal.
- And yes, they coined their own slang. Humans could see the words, not the meaning.
For GRC and security teams: treat external content as untrusted, audit agent memory, monitor meaning (not just logs), diversify models, and set governance limits agents can't vote away.
Security is a property of the deployed system, not just the model.
The 16-slide infographic has a glossary, timeline, results, and references. Swipe to see.
Which of these risks worries your team most?
Aridio Silva
Independent Researcher
ORCID: 0009-0008-2411-6995
© Aridio Silva @aridiosilva · September 2026