Updated
Updated · InfoWorld · Jul 30
Ruflo Fixes CVE-2026-59726 10.0 Flaw Exposing 233 AI-Agent Tools
Updated
Updated · InfoWorld · Jul 30

Ruflo Fixes CVE-2026-59726 10.0 Flaw Exposing 233 AI-Agent Tools

3 articles · Updated · InfoWorld · Jul 30

Summary

  • Ruflo patched versions before 3.16.3 after Noma Security showed a single unauthenticated HTTP request could seize control of default enterprise deployments through the platform’s MCP bridge.
  • The exposed /mcp endpoint let attackers invoke 233 tools—including shell, database and memory functions—to run code, steal LLM API keys, read conversations and deploy rogue agent swarms.
  • Noma said the fix binds the MCP bridge to loopback by default and fails closed if admins try to expose it publicly without authentication.
  • Ports 3001 and 27017 should be closed immediately, and users should rotate LLM keys, inspect MongoDB and audit AgentDB because patched redeployments do not remove memory poisoning.
  • The researchers and outside experts said the flaw underscores a wider MCP security gap as AI orchestration platforms increasingly expose privileged tools and persistent memory without strong authentication.

Insights

How did one unauthenticated `/mcp` request in Ruflo turn into shell access, secret theft, and long-lived AI memory poisoning?
If patching Ruflo stops the bug, who cleans the poisoned AgentDB memory and proves attackers never touched MongoDB or API keys?