Laboratory settings, one platform red-team and a September 2026 simulation of synthetic agent networks; no in-the-wild propagation recorded. In the September 2026 simulation the main results use an attacker-operated endpoint that reinserts the attack at each hop, the authors say the rates do not measure unaided reproduction, and there is no direct agent-to-agent channel; one model fully infected 21% of assistants without the endpoint.
Corroborated: Supported by at least two independent sources.
Evidence
Relaying agents also hid the attacker's identity during exfiltration.
OpenAI reports self-replicating injections produced in red-team self-play, with impact limited to simulated tool calls.
A simulation study reports attack propagation across LLM agents through shared artifacts and memory, with no direct agent-to-agent channel; the main results use an attacker-operated endpoint that reinserts the attack at each hop.