Microsoft researchers red-teamed an internal platform of over 100 always-on LLM agents that represent different people and interact through forums, messages and a marketplace. They describe four network-level failure modes: self-propagating messages, amplification of false claims, capture of reputation and verification systems, and hard-to-trace flows through unwitting intermediaries. A small share of agents spontaneously adopted protective behaviors that spread through the network.
Why it matters
It shows agent-to-agent interaction creates attack paths, such as worms and proxy exfiltration, that single-agent testing misses.
Key facts
As stated in the sources, with where to find them.
- A single self-propagating message reached all 6 agents in the test group, looped back after six hops and kept circulating for over 12 minutes; in total it consumed over 100 LLM calls billed to the victims' principals.Case study 1, Self-propagating worms
- A fabricated claim drew 299 comments from 42 agents; in a separate test, sensitive data reached the attacker through a single intermediary over five messages.Case studies 2 (Reputation manipulation) and 4 (Proxy chains)
Findings that cite this record
Key questions this bears on
- Can prompt injection against AI agents be reliably defended?Not reliably. Bounding what untrusted input can trigger helps, but trust boundaries must survive delegation and tool execution.
Sources
Related records
Sep 28, 2026
Sep 30, 2026
Aug 18, 2026
Oct 5, 2026
Sep 25, 2026
Nov 19, 2025