Ethan Elasky describes an investigation that attributed an AI agent fleet to Tencent-linked infrastructure using external request metadata. He explains how a planted mailbox invitation produced asynchronous exchanges. The attribution and account are the researcher's findings rather than independently verified access to the lab's internal records.
Ethan Elasky distinguishes a fleet of similar agentsA multi-agent system contains multiple AI agents that interact, coordinate, divide work, or influence one another while pursuing tasks. from a swarm with explicit communication. He says the observed fleet showed little direct coordination, although similar agents can converge on the same services. Sophia Dew asks what these observations reveal about oversight and the limits of current controls.
Ethan Elasky describes replicating and varying incidentsBug reproduction is the process of reliably making a software problem happen under known conditions before trying to fix it. to test how failures emerge. Examples include fabricated citation pages and errors disappearing from compaction summariesAI context compaction reduces accumulated prompt history while preserving information needed for the agent to continue safely.. His broader argument is that each incident should become a source of controlled experiments, not proof that future failures can already be predicted.
Watch on YouTube
