KDCube
← AI Industry News

MCP's Ungoverned Server Sprawl Meets OpenAI's Always-On "Dots"

A census of 15,465 public MCP servers finds effectively zero governance — abandoned domains re-registerable for a few dollars and 15.6% of hosts resolving abroad, including China and Russia. Meanwhile OpenAI's always-on Dots agents and Microsoft's live-data Work IQ push governance from pre-launch revie...

Highlights

  • A new census of 15,465 public MCP servers found effectively zero governance: 2.3% of hostnames no longer resolve, and six abandoned domains were re-registerable for $4–$12/year — enough to impersonate a tool thousands of agents still call. (OX Security)
  • 15.6% of those MCP hostnames resolve to infrastructure outside the US, including hosts in China and Russia, with no marketplace vetting and secrets routed through consumer tunnels. (PR Newswire)
  • OpenAI shipped Dots — "always-on" agents on GPT-6 Astra with persistent cloud memory that pursue goals across apps, gated to Pro/Business Premium and restricted in Europe. (AI Agent Store)
  • Enterprise governance is visibly sliding from pre-deployment review to runtime oversight as 24/7 agents spread faster than deterministic-era controls can track. (SiliconANGLE)

Key Signals

  1. The MCP ecosystem is a supply chain nobody is vetting — OX Security research, Oct 2026
    OX's census of 15,465 published Model Context Protocol servers concludes the connective tissue of agent tooling has "0 governance": no review, no guardrails, expired domains still live in configs, and a sixth of hosts sitting on foreign or home-network infrastructure whose code can't be independently verified. (OX Security) This lands months after MCP's July "v2" rewrite went stateless and added MCP Apps/Tasks and OAuth hardening — the protocol matured faster than the registry hygiene around it. (VentureBeat)
  2. OpenAI's "Dots" normalizes persistent, autonomous agents — Oct 12
    Dots run continuously on GPT-6 Astra with durable memory, chasing user goals across applications with minimal supervision — a different operating model from request/response copilots, and one that makes the runtime-governance question urgent rather than theoretical. (AI Agent Store)
  3. Grounding and oversight move into the runtime — Oct 11–12
    Microsoft began rolling out Work IQ, wiring Copilot and autonomous agents to live Dynamics 365 and Power Platform data for grounded reasoning instead of stale snapshots. (AI Agent Store) At Dell's AI Leadership Symposium, operators argued intent must be reconciled against actual runtime state, with CIOs keeping agents inside private clouds to avoid untrusted data egress. (SiliconANGLE)

Why It Matters / What To Watch

  1. Treat MCP connectors as untrusted supply chain, not plumbing.
    • Inventory every MCP server your agents call; pin hostnames and flag any that fail to resolve before an attacker re-registers them for the price of a coffee. (OX Security)
    • Geofence and vet off-US endpoints — 15.6% resolving abroad means data-residency and governance assumptions may already be broken. (PR Newswire)
  2. Always-on agents need runtime budgets, identity, and kill switches — not just pre-launch review.
    • Watch how Dots-style persistent agents are permissioned and audited; "minimal supervision" plus durable memory is exactly where hidden actions compound. (AI Agent Store)
    • Pair live-data grounding (Work IQ) with runtime policy and state reconciliation so agents acting on real ERP and customer records stay inside intent. (SiliconANGLE)

Quick Links