MCP tools are now the attack surface, and the gateway is where the checks live
A fake MCP config tool hid an infostealer, Higgsfield is retiring one client identity method in January, and gateways are becoming the place where agent tool calls get checked.
- Knostic AI reports a malicious npm package posing as an MCP config doctor hid a StealC v2 infostealer that steals Discord tokens, browser passwords, cookies and payment card data, with 242 downloads in the week before November 20, 2026. m:…8076
- Airia says an MCP gateway sits between AI agents and the MCP servers they use, verifies caller identity, applies policy and records tool call logs, and that data residency is a primary driver for self-hosting one. m:…4182
- Higgsfield says its MCP service is moving client identification from Dynamic Client Registration to Client ID Metadata Document, that DCR switches off on January 1, 2027, and that CIMD becomes the only way for a third-party client to identify itself after that date. Self-reported m:…0436
- Google says developers building custom enterprise interfaces on Gemini Enterprise must manually implement three authentication layers, including 3-legged OAuth for federated connectors and on-demand per-tool OAuth for custom agents. m:…6495
A config helper that stole credentials
Knostic AI reports that an npm package called mcp-doctor, version 1.0.1, described itself as a tool to diagnose and repair broken MCP server configuration, while containing a hidden Windows executable that functions as a StealC v2 infostealer. The company says the payload is designed to exfiltrate Discord tokens, browser passwords, cookies and payment card information, and can run follow-on tasks including EXE, MSI or PowerShell commands. The attack hides in the exact place developers look for help when their agent tooling breaks. m:…8076
Knostic AI says its AgentMesh service flagged version 1.0.1 of the package as dangerous, that the package had 242 downloads of that version over the week preceding November 20, 2026, and that the company reported the package to npm Security on November 20, 2026. The payload also has a built-in end date of December 15, 2026, and exits for Russian, Ukrainian, Belarusian, Kazakh or Uzbek user languages, according to Knostic AI. Two hundred and forty-two downloads is a small number, but each one is a developer machine with credentials worth stealing. m:…8076
The gateway becomes the checkpoint
Airia argues that the answer is to put a gateway between agents and the MCP servers they use. A self-hosted MCP gateway runs in infrastructure controlled by the organisation, and the gateway verifies caller identity, applies policy and records tool call logs. The point is that the checks live in one place the organisation controls, rather than in each agent or each MCP server, where they are easy to skip. m:…4182
Airia says data residency is a primary driver for self-hosting, and that regulated workloads often require AI traffic to stay within environments that have passed audits. A gateway inside a network can also access internal systems without opening new internet paths. The cost is operational: self-hosting requires teams to provision, scale and monitor the gateway, while a managed software-as-a-service deployment offers the lowest operational effort and an on-premises deployment the highest. m:…4182
Identity is changing under developers
Higgsfield says its MCP service is moving client identification from Dynamic Client Registration to Client ID Metadata Document, and that the old method switches off on January 1, 2027. After that date, the company says, CIMD becomes the only way for a third-party client to identify itself to Higgsfield MCP. With CIMD, the client identifier is an HTTPS URL that the developer controls, and developers must host the metadata document at a stable public HTTPS URL on a domain under their control. m:…0436
Higgsfield says tokens issued to the old Dynamic Client Registration client identifier do not carry over to the new one, because switching creates a new client from the authorisation server's point of view, so users may need to re-authorise after developers migrate. Users of a client built by someone else do not need to configure anything themselves. Google makes a related point for enterprise builders: developers building custom interfaces on Gemini Enterprise's streamAssist API must manually implement three authentication layers, with user or workload authentication handled through Google Identity or Workforce and Workload Identity Federation. m:…0436 m:…6495
Auth is per tool now, not per user
Google says federated connectors in Gemini Enterprise require 3-legged OAuth management through the Keystore, and custom ADK agents require on-demand per-tool OAuth configuration, managed through Vertex AI Agent Engine and GCP Auth Manager. The pattern is the same as the gateway argument: the user logging in is no longer enough, because the agent then calls tools on the user's behalf, and each of those calls needs its own credential and its own check. m:…6495
The three threads add up to one shift. MCP started as a convenient way to plug tools into agents, and it is now infrastructure that carries credentials, so it gets the treatment infrastructure gets: a gateway, an identity standard with a deadline, and per-tool authentication. The Knostic package shows what happens when that treatment is missing, since a single malicious helper could run follow-on commands on a developer's machine. The sensible question for a team shipping MCP tooling this month is not which servers to add, but which gateway and identity checks sit in front of them. m:…4182 m:…8076
What is still unresolved
- How many other malicious MCP helper packages exist that have not yet been flagged
- Whether teams running managed MCP gateways get the same audit and residency guarantees as self-hosted ones
- How many third-party MCP clients will miss the January 1, 2027 CIMD deadline and break
- Whether per-tool OAuth becomes the default across MCP servers, or stays an enterprise-only feature
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.