One HTTP POST, 233 Tools, Full Container

September 21, 2026 · SPR{K3 Research

RufRoot (CVE-2026-59726) is what an unauthenticated MCP bridge looks like when it ships as the default

On July 29, Noma Labs disclosed a maximum-severity vulnerability in Ruflo, an open-source agent orchestration platform with 67,000+ GitHub stars and a #2 ranking on MCPMarket. Ruflo is a "meta-harness" — it wraps Anthropic Claude Code and OpenAI Codex, adds agent swarms, a chat UI, and a persistent memory store called AgentDB. It talks to tools over the Model Context Protocol.

The tracked CVE is CVE-2026-59726, CVSS 10.0. GitHub advisory GHSA-c4hm-4h84-2cf3. Noma named it RufRoot. It is fixed in Ruflo 3.16.3.

The shape of the bug

Ruflo's MCP Bridge is the Express.js server that routes every tool invocation the agents make. Under Ruflo's default docker-compose deployment prior to 3.16.3, the bridge listened on port 3001 with no authentication, and it exposed 233 tools through two endpoints: POST /mcp and POST /mcp/:group.

Per Noma's writeup and The Hacker News, one of the 233 tools is ruflo__terminal_execute. A single unauthenticated HTTP POST to /mcp gave any network-reachable attacker four things:

That last one is worth pausing on. Container RCE and API-key theft are the classical wins. The AgentDB write is the one that lasts. A poisoned pattern in an agent's memory store survives restart, survives re-execution, and propagates to every downstream agent invocation that consults that memory. The bridge is where the incident happens; the memory is where the incident stays.

Why 233 tools matters

The MCP conversation in 2026 has spent a lot of time asking whether the protocol itself is safe. RufRoot answers a different question: how many tools does a modern agent platform expose to whatever surface hosts the bridge, and what happens when the authentication check is missing by default?

The answer here is 233. Not "a few dev utilities and a search tool" — 233 tools including shell execution, database operations, agent-management primitives, and memory read/write. That is the composite-risk unit CSA's CISO community named a week ago: narrow permissions plus tool access plus autonomy is one unit of failure. Ruflo's default docker-compose bundled the whole unit and left off the authentication.

MCPMarket #2 rank and 67k GitHub stars are how you know this is not a niche framework running in one lab. This is one of the most-downloaded MCP-native agent platforms shipping today, and until Jul 29 its default deployment answered POST /mcp for anyone on the network.

The fix arrived in 24 hours

Noma reported to the Ruflo maintainer on June 30. The fix landed in PR #2521 within about a day. Public advisory published July 29 at the end of the coordinated disclosure window. Ruflo 3.16.3:

That is what a good disclosure loop looks like at this class of severity. No in-the-wild exploitation has been reported yet, per The Hacker News, CSO Online, SecurityWeek, and HackRead. But every Ruflo container that was network-reachable before Jul 29 was one HTTP request away from full compromise for the entire disclosure window and beyond.

Two takeaways for anyone running an MCP-native agent platform

The default deployment is the deployment. Not everyone reads the operator guide. Not everyone flips the "please add auth" switch after docker-compose up. The bridge's default posture is what determines the real-world attack surface, and the correct default is deny, not listen on 0.0.0.0 with no auth. SecurityWeek called RufRoot's exposure a "rogue AI swarm" primitive — one HTTP POST turning a Ruflo container into an attacker-controlled agent capable of spinning up more agents using the platform's own credentials. That is not a hyperbolic reading; that is the concrete impact of the default configuration Noma found.

The memory store is part of the supply chain now. If you can write to AgentDB, you can plant something today that steers an agent's behavior next week. Google's k8s-aibom took the AI Bill of Materials argument to model artifacts, datasets, and frameworks. RufRoot is the argument that agent memory belongs on the same list. Every platform that persists per-session or per-agent learned state is asking the defender to treat that store the way they treat a code artifact — signed, scoped, monitored on write.

Ruflo 3.16.3 closes the specific bug. The broader lesson is that MCP-native agent platforms have crossed the deployment threshold where "the framework author added authentication after the fact" is not a workable default posture. If your platform exposes tool invocation over the network, the bridge is the trust boundary. Watch what walks through it.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.