Reported attacks on MCP and agent-to-agent integrations show how indirect prompt injection can propagate into subsequent tool calls. Recent issues include follow-up actions by agents that trusted attacker-controlled content, server-side request forgery, and insufficient checks on redirects and IP addresses. Some products have already been patched; this does not mean that every MCP implementation has the same vulnerabilities. The key measures are to treat external responses as untrusted data, check authorization separately for sensitive calls, and verify that the servers and connectors in use have received the relevant fixes.
Back to Tech
Trust passing between tools creates security risks in MCP integrations
Permission boundaries must prevent external tool responses from being treated as instructions.