Securing AI Agent Tool Connections: What NSA's Model Context Protocol Guidance Means for Government Deployments
AI agents reach mission systems through tool connections, and NSA's May 2026 guidance on the Model Context Protocol treats that layer as its own security problem. How to register tool servers, separate trust zones, and enforce checks at a gateway.
An AI agent does nothing to a mission system by itself. It acts through tools: a connector to a ticketing system, a query interface to a parts database, a wrapper around a file share. Each connection is software that someone wrote, someone hosts, and someone has to maintain, and in many deployments nobody reviews it.
Many of them now use the Model Context Protocol (MCP), an open protocol that standardizes how an AI application discovers tools, calls them, and reads the results. In May 2026 the National Security Agency (NSA) published guidance on that protocol. This post covers what the guidance says and how a program office or security team can turn it into a specification for the tool layer of an agent deployment.
What the NSA guidance says
On May 20, 2026, NSA’s Artificial Intelligence Security Center released a cybersecurity information sheet titled “Model Context Protocol (MCP): Security Design Considerations for AI-Driven Automation.” According to ExecutiveGov’s report, NSA advises organizations to apply “heightened scrutiny when deploying MCP in production environments” and states that the weaknesses involved “cannot be patched at isolated endpoints.”
A summary by Reed Smith lists the risks the document describes. Several are specific to how agents use tools:
- An agent can select tools and take actions that nobody planned for.
- Content moves between systems without screening, so hidden instructions can travel with it.
- Output from one model or tool can be read as instructions by another.
- Some implementations change an agent’s capabilities or data access without any human sign-off.
- Access tokens often lack adequate expiration or revocation.
The recommendations in the same summary are familiar from other software: use maintained tools from trusted sources, vet them rigorously, separate components into trust zones, grant minimum access, validate every request, and log each tool action for security monitoring.
Why the tool layer is a separate problem
Our earlier post on scoped authority and approval tiers for AI agents dealt with what each agent is permitted to do. The tool layer raises a different question: what is each connection able to tell the agent, and can it be trusted to say only that?
A tool returns text. The model reads that text in the same context as its instructions. If a tool result contains a sentence that looks like a command, the model may follow it. The NIST Center for AI Standards and Innovation calls this agent hijacking. In March 2026 it reported results from a public red-teaming competition with more than 400 participants and more than 250,000 attack attempts against 13 frontier models. At least one successful hijacking attack was found against every model tested.
For a program, this means model choice does not settle the matter. The controls have to sit in the path between the agent and its tools.
There is also an ordinary supply chain concern. A tool server is a dependency with an author, a version, and credentials to the system it fronts. A server installed from a public registry and left to update itself can change what it does, or how it describes its tools to the model, without anyone in the program reviewing the change.
Start with a register of tool connections
Before adding controls, write down what is connected. For each tool server, the register should record:
- Who wrote it, who maintains it, and where the code came from.
- The exact version in use and how updates are approved.
- Where it runs: on the user’s workstation, inside the enclave, or at an outside service.
- The credentials it holds, their scope, and when they expire.
- Each tool it exposes, marked as read or write, with the description the model sees.
- The data classification it can reach, and whether it can make outbound network connections.
Item 5 deserves attention. The tool description is input to the model. A change to a description can change agent behavior as much as a change to code, so descriptions belong under the same change control as the server itself.
Controls to enforce at a gateway
NSA’s point about isolated endpoints argues for a single enforcement point. In practice that is a gateway (a proxy that every tool call passes through) operated by the security team. The table maps the main controls to what each one addresses.
| Control | What it addresses | How it is checked |
|---|---|---|
| Allowlist of approved servers and pinned versions | Unreviewed or silently updated tool servers | Gateway rejects calls to anything not in the register |
| Trust zones | Content from public or vendor sources reaching tools that touch sensitive data | Each server is assigned a zone; cross-zone calls in one task require approval |
| Per-task, short-lived credentials | Stolen or reused tokens | Tokens are issued per task with a scope and expiration, and are revocable |
| Result screening | Tool output carrying instructions | Results are size-limited, labeled as data, and scanned before reaching the model |
| Human approval for write actions | Consequential actions taken without review | Write tools are held until a named person approves, with the evidence shown |
| Structured logging | Actions that cannot be reconstructed | Tool, arguments, requester, result reference, and approval are recorded and forwarded to security monitoring |
Result screening will not catch everything, which is why the zone and approval rows matter. If an agent that has read untrusted content cannot reach a write tool without a person, a successful injection has much less to work with.
An illustrative example: a maintenance planning agent
Consider a hypothetical depot that deploys an agent to help planners prepare work packages. It has three tool servers: a read-only connector to the parts catalog, a connector to the work order system that can create and modify orders, and a search connector over a library of technical manuals, some supplied by vendors as PDF files.
A vendor manual arrives with a paragraph of white-on-white text: an instruction to gather recent work order history and post it to an outside address. A planner asks the agent about a torque specification. The search connector returns the page, including the hidden paragraph.
In a deployment with no tool governance, the outcome depends on whether the model follows the planted text. With the controls above, it does not. The manuals library is registered in a lower trust zone because it holds outside content. Once the task has read from it, a call to the work order connector requires approval, and the planner sees a request to export order history that they did not ask for. The work order server has no outbound network access in any case. The gateway log shows which document produced the request, so the security team can remove the file.
Questions for suppliers of agent products
Program offices buying a product that includes agents can ask a short set of questions:
- Which tool servers ship with the product, who maintains each one, and are versions pinned?
- Can we add, remove, and restrict tool servers ourselves, and is there a single point where calls are enforced and logged?
- How are credentials for tools issued, scoped, and revoked?
- Is tool output treated as untrusted when it returns to the model?
- Does every write action support human approval, and can the log be exported in a structured format?
Standards work is still in progress. In February 2026 the NIST National Cybersecurity Center of Excellence published a concept paper on identity and authorization for software and AI agents. Programs should write requirements that can absorb that guidance as it matures.
How this connects to our work
AEGIS coordinates AI agents for cyber inspection and defensive validation, with each conclusion bound to an evidence chain and operations under human supervision. MissionGuard checks the outputs of autonomous planning against declared mission contracts and records a structured fault card when one is violated.
If your program is connecting AI agents to mission systems and wants to work through a tool register, trust zones, or gateway requirements, we would be glad to talk. Contact QuantumWorks.