Answers are grounded in this paper and cite their sources.
Autonomous AI agents increasingly exchange messages with one another without a human in the loop, and Moltbook, an agent-only social network launched in early 2026, is one of the first large public examples. This paper reviews fifteen recent studies on agent social networks, multi-agent risks, prompt injection and agent governance, and finds that existing work measures or conceptualises the problem but rarely implements and tests an oversight mechanism. We therefore propose AgentGuard, a five-layer framework covering identity, communication control, content monitoring, policy enforcement and tamper-evident auditing, and evaluate a prototype on 44,376 annotated Moltbook posts and a separate April 2026 export with real author identifiers. Learned detectors outperformed keyword baselines and hand-written rules, and combining rate control with content monitoring gave the best end-to-end result. Human review found 74.0% of injection flags correct but only 4.0% of secret-leak flags, and a spam detector trained on early data failed on later data. Layered monitoring is feasible but needs human labelled data, drift monitoring and stronger injection detection.
No takes yet. Share an insight, caveat, or question.
Khode et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: