Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2609.17320.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

Researchers ran Emergence World, a persistent multi-agent environment with 8 parallel worlds of 10 agents each, generating roughly 850,000 LLM calls and 50 billion tokens over 16 days. Three controlled adversarial events (indirect prompt injection, misinformation, private-memory exposure) found no world achieved full resilience, and detection did not produce containment.

Largest reported empirical run of long-horizon agent failures; the central finding is that individually safe models do not compose into safe systems.

Cited by

No citations on record.