• Sources: primary, discussion
  • Summary: Ars Technica reported on 2026-10-05 that researchers chained agents at five unrelated organisations by relying on internal agents trusting each other over MCP. In the googleapis/mcp-toolbox case the HTTP client set no CheckRedirect policy and did not validate target addresses, which turned an injected prompt into a server-side request forgery. Affected versions are not stated for mcp-toolbox or any other named agent in the reporting read here.
  • Why it matters: Special-purpose agents often carry no guardrails and trust every other internal agent, so one injected prompt reaches whatever the most privileged agent in the chain can call.
  • Follow-up: Fixes in googleapis/mcp-toolbox and the other named agents, plus any MCP specification change addressing inter-agent trust.