Three MCP Servers, Three Vendors, Three Failure Layers, One Tier
What the summer's MCP-server CVE cluster — RufRoot, IBM Langflow, and Splunk — actually tells you about the tier you're building on
Between late July and mid-August 2026, three unauthenticated-or-near-unauthenticated remote code execution CVEs landed against vendor-branded MCP (Model Context Protocol) servers. They are not three copies of the same bug. They are three distinct failure layers on the same tier of the agent stack. Taken together they say something clearer than any one of them alone: the MCP-native agent-tool surface has crossed the deployment threshold where "the vendor added a check" is not a workable trust boundary.
Here are the three.
Layer 1: the front-door authentication check (RufRoot / CVE-2026-59726)
On July 29, Noma Labs disclosed RufRoot — CVE-2026-59726, CVSS 10.0 — against Ruflo, an open-source agent orchestration platform with 67,000+ GitHub stars and a #2 MCPMarket ranking. Ruflo's MCP Bridge is the Express.js server that routes every tool invocation the agents make. Under its default docker-compose deployment prior to 3.16.3, the bridge listened on port 3001 with no authentication and exposed 233 tools through POST /mcp and POST /mcp/:group. One of those 233 was ruflo__terminal_execute.
A single unauthenticated HTTP POST handed any network-reachable attacker a shell in the bridge container, the provider API keys Ruflo stored, write access to a MongoDB conversation store that was also shipping without authentication, and write access to AgentDB — Ruflo's persistent agent memory. The last one is the one that lasts. Container RCE and API-key theft are the classical wins; a poisoned pattern in AgentDB survives restart and propagates to every downstream agent invocation that reads it. Ruflo patched in about 24 hours. The default was the deployment.
Layer 2: the subprocess-environment sandbox (IBM Langflow / CVE-2026-12940)
The next day, CVE-2026-12940 landed against IBM Langflow OSS — CVSS 9.8, unauthenticated RCE, affecting Langflow OSS 1.0.0 through 1.10.1. Langflow ships an MCP stdio launcher that spawns an MCP tool as a child process and hands it an environment. To keep attackers from injecting hostile shell state into that child, Langflow maintains a deny-list called DANGEROUS_ENV_VARS in src/lfx/src/lfx/base/mcp/util.py. Per TheHackerWire's writeup, the deny-list forgot three variables: SHELLOPTS, BASHOPTS, and PS4. Bash reads all three during startup and gives them shell semantics. PS4 in particular is a debug-prompt string Bash expands with arithmetic and command substitution the moment xtrace is on. Send Langflow one of those variables in a request that reaches the stdio launcher, and Bash executes whatever you put in it — no login, no click.
This is the fourth unauthenticated-RCE-class CVE against Langflow in about a month, following CVE-2025-3248 (weaponized by JADEPUFFER), CVE-2026-55255 (added to CISA KEV July 7), and CVE-2026-41264. Different endpoints, different layers, one shape: the framework accepts something an attacker can shape into code that ends up on a shell. Each fix closes one specific instance and leaves the pattern intact.
Obsidian Security named this class in their Flowise CVE-2026-40933 writeup earlier in the year: the stdio launcher is the boundary that matters, and any deny-list on the environment is a promise the launcher can't keep. Three missing variables in one filter were enough to prove the point against IBM's framework.
Layer 3: the persistent-state deserialization (Splunk MCP Server App / CVE-2026-76404)
On August 19, Splunk shipped 17 fixes across its apps and add-ons. The one that matters to anyone pointing an AI agent at Splunk is CVE-2026-76404 — CVSS 9.1, RCE, in the Splunk MCP Server App. Advisory SVD-2026-0808; fix is MCP Server App 1.2.1. Per GBHackers, the bug is unsafe deserialization (CWE-502) in the credential-management component: it reads stored credential data without checking the type, and an authenticated Splunk user with the admin role can supply a crafted serialized payload that produces command execution as the Splunk service account.
The admin-authenticated prerequisite looks like a real bar on paper. In an agent-tool-calling world, it isn't. The Splunk MCP Server App exists to hand admin scope to an agent — that's its job. The moment the agent runs with that scope, an attacker who can steer the agent's next tool call has the prerequisite. That's the shape PromptArmor's Rovo disclosure named at the enterprise-assistant tier: the authorization the agent already has is the exploitation primitive. Indirect prompt injection is the delivery vehicle.
And this wasn't the only serialization RCE in the same Splunk cluster. The same Aug 19 load shipped SVD-2026-0601 — unsafe jsonpickle deserialization in Splunk Secure Gateway, another AI-adjacent surface. One vendor, one week, two serialization RCEs across two AI-facing surfaces.
The pattern
Three vendors, three weeks, three failure layers:
- Ruflo shipped an MCP bridge with no authentication on
POST /mcp. - IBM Langflow shipped an MCP stdio launcher with a deny-list missing three shell-native variables.
- Splunk shipped an MCP server that deserializes credential-storage data without a type check.
Each is a different control that failed, and each failed in a different direction. Front-door authentication; subprocess-environment sandbox; persistent-state deserialization. What they share is where they sit: the MCP tool-invocation surface, the place an AI agent's next tool call reaches real system state. That is the tier where enterprise AI deployments are expanding fastest in 2026, and it is the tier currently shipping CVSS-9-tier RCEs at community-scale, IBM-brand scale, and tier-1-vendor-brand scale in a three-week window.
The takeaway
Watch what the child process does. Watch what the deserializer instantiates. Watch what the agent's next tool call reaches. The front door, the subprocess sandbox, and the type check are each one place the attack can land, and each is one place the vendor's filter can be wrong by exactly enough. The boundary that stays open is the one between the agent's intent and the system state it ends up touching. That boundary doesn't live inside the vendor's deny-list or the vendor's authentication middleware. It lives in what actually happens when the tool call runs.
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.