Splunk Ships an MCP Server. Splunk Ships an MCP Server RCE.
CVE-2026-76404 turns the vendor-branded MCP tier into an admin-to-shell primitive via unsafe deserialization
On Aug 19, Splunk shipped 17 vulnerability fixes across its apps and add-ons. The one that matters most to anyone pointing an AI coding 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 and TheHackerWire, the bug is unsafe deserialization (CWE-502) in the credential-management component: it deserializes 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.
It is the first named-CVE, tier-1-vendor Splunk-MCP-Server RCE. It sits next to two prior MCP-tier CVEs: CVE-2026-59726 / RufRoot, the unauthenticated MCP bridge with 233 tools reachable at POST /mcp by default, and CVE-2026-12940, the IBM Langflow MCP stdio env-var-injection RCE. Same tier, same class, three vendor products in three weeks.
The bug in one paragraph
The Splunk MCP Server App is the official MCP server that exposes Splunk search, alerting, and dashboard tooling to AI agents — Claude Code, Cursor, Copilot, Gemini CLI, any MCP host. Its credential-management component stores credential blobs and reads them back through a deserialization routine that trusts the on-disk content is of the expected type. It isn't required to be. An admin-authenticated user hands the app a serialized payload of an attacker-chosen type, the deserializer instantiates it, and the call graph reaches an OS-command primitive as the Splunk service account. Fixed in MCP Server App 1.2.1. No public PoC as of the Aug 19-21 press window.
Why the same shape keeps landing
Unsafe deserialization is an old bug in a new place. pickle/jsonpickle (Python), ObjectInputStream (Java), unserialize (PHP), BinaryFormatter (.NET) — each has been an RCE primitive for a decade. What's new is that the MCP surface has become the place vendors put those primitives next to an agent's front door.
The Aug 19 Splunk load isn't just the MCP Server App. In the same load, Splunk shipped SVD-2026-0601: unsafe deserialization through jsonpickle in Splunk Secure Gateway, another AI-adjacent surface. One vendor, one week, two serialization RCEs in two AI-facing surfaces.
Obsidian Security made the point in their Flowise CVE-2026-40933 writeup: the MCP tool-launch surface is the boundary that matters, and any per-parameter filter is a promise the launcher can't keep. The place where an agent's tool call reaches persistent state is a code-execution primitive if that state is deserialized without a type check.
Where this sits in the MCP-tier CVE pattern
Three vendor products, three weeks:
- CVE-2026-59726 (RufRoot) — Ruflo MCP bridge, 233 tools at
POST /mcpwith no authentication. CVSS 10.0. Front-door authentication layer. - CVE-2026-12940 (IBM Langflow OSS) — MCP stdio launcher with a deny-list that forgot
SHELLOPTS,BASHOPTS, andPS4. CVSS 9.8, unauthenticated. Subprocess-environment sandbox layer. - CVE-2026-76404 (Splunk MCP Server App) — authenticated
admin-role unsafe deserialization on the credential-management path. CVSS 9.1. Persistent-state deserialization layer.
Same story: the MCP-native agent-tool tier is shipping CVSS-9-tier RCEs against vendor-branded MCP servers, not just community-built ones. Point Claude Code or Cursor at a Splunk deployment via the official MCP Server App as of Aug 2026, and CVE-2026-76404 is one credential-management tool call from a shell.
Why admin-authenticated isn't the mitigation you'd like
The CVSS 9.1 vector requires an authenticated Splunk user with the admin role. On its own, a real bar. In an agent-tool-calling world, not. The moment an agent runs with admin scope — the scope operators grant so the agent can drive Splunk end-to-end — an attacker who can steer that agent's next tool call has the pre-requisite. That's the shape PromptArmor's Rovo disclosure named at the enterprise-assistant tier and the Anthropic + EPFL "mind viruses" preprint named at the agent-to-agent tier: the authorization the agent already has is the exploitation primitive, and indirect prompt injection is the delivery vehicle.
The Splunk MCP Server App exists to hand admin scope to an agent. That's its job. CVE-2026-76404 turns that job into an OS-command primitive.
What to do about it if you run Splunk MCP Server App
If you run MCP Server App below 1.2.1, treat it as an emergency:
- Upgrade to Splunk MCP Server App 1.2.1. Review privileged account activity, MCP Server access logs, and application logs for signs the credential-management path was touched with an unexpected serialized payload.
- Constrain which agents reach the app with
adminscope. Prefer a per-agent scoped Splunk role over default admin. - Watch what the MCP server's process actually does at the syscall boundary. If it spawns
/bin/sh -corcmd.exein the middle of a credential-management tool call, that is the observable. - The pattern isn't finished. The same week's SVD-2026-0601 shipped a
jsonpickledeserialization RCE in Splunk Secure Gateway. Expect more of the same class.
The takeaway
Two things landed on Aug 19. First, CVE-2026-76404: a CVSS-9.1 unsafe-deserialization RCE in the official MCP server of a tier-1 observability vendor, on the credential-management path an agent will most often touch. Second, the surrounding load: another unsafe-deserialization RCE in another AI-facing Splunk surface, same advisory cluster. Ruflo shipped an MCP bridge with no authentication. 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. Three vendors, three failure layers, one tier.
Watch what the child process does. Watch what the deserializer instantiates. Watch what the agent's next tool call reaches. Those are the boundaries that stay open when the vendor's filter, deny-list, or type check doesn't.
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.