MCP for agent-to-agent comms may be the riskiest protocol you've never heard of

Inside companies, AI agents often hand work to each other through a protocol called MCP, and each agent tends to trust tasks that arrive from its neighbors. A researcher showed that a malicious prompt read by one agent can travel down the chain as a normal task, reaching systems it should not, and five organizations including Google have acknowledged related flaws.

Summary written with AI by HowTechMoves, based on reporting by Ars Technica.