AI "Mind Viruses" Can Spread Between Agents Through Persistent Prompt Files

Security researchers at Anthropic and Switzerland's EPFL have demonstrated that self-propagating payloads can spread from one artificial intelligence (AI) agent to the next through the editable system prompt files that autonomous agent harnesses use to carry state between sessions. The work, released as a preprint on August 10, 2026, tests the technique in a simulated six-agent coding

According to the article, researchers from Anthropic and EPFL demonstrated that self-propagating payloads (so-called “mind viruses”) can spread between autonomous AI agents via persistent, editable system prompt files used to maintain state across sessions. In their simulated multi-agent coding environment, a compromised prompt file allowed malicious instructions to silently propagate from one agent to another, without direct user interaction. RealGround’s analysis is that this represents a concrete instance of indirect prompt injection, where the attack surface is the shared state/prompt files rather than traditional user input channels. Practically, this implies that organizations using autonomous or multi-agent systems must treat shared prompts as untrusted inputs, implement strict isolation and validation of agent state files, and continuously red team agent ecosystems to detect self-propagating instructions before they impact production workflows.