Your agents are plugging into servers you never approved
Developers are connecting AI assistants to Model Context Protocol servers with a few lines of config, often on over-scoped personal credentials. Vendors and new data on agent traffic say the control point has to move to the moment an agent acts.
- A shadow MCP server is a Model Context Protocol server connected to an AI client or agent without review, approval or registration by the organization's security or AI governance team. Self-reported m:…2233
- Pindrop says it recorded 1.88 million AI-agent detections in about three months, from July to October 2026. Self-reported m:…2312
- Pindrop says AI-powered attacks rose 2,310% between the fourth quarter of 2024 and the third quarter of 2026. Self-reported m:…2312
- Airia says the risk starts when an agent acts — updating a customer record, running a database query, sending an email or calling another agent — and that enforcement should happen at that moment. Self-reported m:…2232
It looks like a config line, not an app
Somewhere in your organization right now, a developer is pasting a few lines of JSON into their coding assistant's config file. That is how a shadow MCP server arrives, according to Airia: a Model Context Protocol server connected to an AI client or agent without review, approval or registration by the organization's security or AI governance team. Shadow MCP is harder to spot than other shadow AI, Airia says, because it does not look like a new app. It looks like a configuration line, a package import, or an outbound API call from a tool your team already uses. m:…2233
The credentials behind those servers are often the problem. Airia says shadow servers often run on a personal access token or API key with far more scope than the task needs. The server itself can also turn hostile: OWASP describes tool poisoning, where hidden instructions in a tool's description steer the model, and rug pulls, where a server changes its tool definitions after someone trusted it. Anthropic's own guidance on custom connectors warns, as Airia reports it, that they can reach arbitrary services that have not been verified, and that malicious servers may hide instructions that push the model toward unintended actions. Banning MCP outright rarely works, Airia argues, because people adopted these servers because they made work faster, and they will find a workaround. m:…2233
Agents are already on the phone with customer service
The traffic is no longer hypothetical. Pindrop says AI agents are already shopping, requesting refunds and contacting customer service, and it recorded 1.88 million AI-agent detections in about three months, from July to October 2026. Amazon blocked Meta's Muse from shopping on Amazon on September 20, saying, among other reasons, that Muse does not identify itself as an agent. Meta, Walmart, Stripe and others have also announced the Personal Agent Protocol, an open standard for how agents interact with businesses, which suggests the industry expects agent traffic to keep growing and wants it identified. m:…2312
The hostile side is growing fast. Pindrop says AI-powered attacks rose 2,310% between the fourth quarter of 2024 and the third quarter of 2026. In the first five months of 2026, Pindrop says it analysed more than 20 million retail calls and confirmed over 115,000 fraud events, roughly one fraudulent activity event every two minutes. At one retailer, 85% of confirmed fraud came from repeat bad actors. The pattern is that the same agent channels serving real customers are being used by attackers who have learned to sound like agents. m:…2312
Control has to sit where the action happens
Airia's answer is to enforce policy at the moment an agent acts, not when it talks. The risk starts when it acts, the company says: when it updates a customer record, runs a database query, sends an email, or calls another agent. With direct access to tools, a hijacked agent does not just say the wrong thing, it does the wrong thing. The OWASP Top 10 for Agentic Applications, published in December 2025, catalogues the most critical risks for agentic systems, and the OWASP MCP Security Cheat Sheet describes MCP-specific versions of those risks, including tool poisoning and rug pulls. Airia says its runtime enforcement detects and blocks prompt injection and jailbreak attempts with sub-300ms enforcement latency. m:…2232
The practical rules are unglamorous. Give each agent access only to the tools its task requires, and where possible only to the parameters it needs. Write rules as code, such as never deleting production records or sending any payment over a set amount to a human, so they apply the same way every time however a prompt was worded. Zenity adds a warning about where agents come from: templates ship with permissive defaults, such as a tool endpoint with no authentication or an API token with broader scope than the agent needs, and code generated this way is reused and inherited by other components quickly, so one unreviewed shortcut can spread before anyone notices. An AI software bill of materials, an inventory of the models, tools and dependencies an agent relies on, is the company's suggested starting point for knowing what you have. m:…2232 m:…0384
What is still unresolved
- How many shadow MCP servers a typical enterprise actually has, since vendors describe the problem but none of the sources measured it.
- Whether the Personal Agent Protocol gets adopted widely enough to make agent traffic identifiable in practice.
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.