The real workflow
Where MCP-Connected AI Assistants enters the work
Agency teams may connect several client mailboxes, drives, knowledge sources, and project systems to a common assistant workflow.
MCP-connected assistants can discover resources and call tools exposed by local or remote servers, creating a reusable bridge between AI and files, APIs, databases, commands, and business systems.
Tool descriptions, resource content, server responses, and resumed session events can carry malicious instructions.
MCP is a protocol, not a security guarantee. The effective boundary depends on the client, server implementation, transport, scopes, tokens, local process privileges, consent, and downstream systems.
The presence of this path does not prove an incident. It identifies the boundary that should be checked before more sensitive context or authority is added.
Tool-specific boundary
Inspect the real access points.
What may carry context
MCP resources and prompts
local stdio server processes
remote tools, OAuth scopes, APIs, and downstream services
Settings to verify
Server origin, command, and transport
OAuth scopes, token audience, and consent
Filesystem, network, session, logging, and downstream permissions
Why this context matters
The consequence for agencies
Agency risk compounds when staff, contractors, shared tools, and reused credentials create paths between otherwise separate client environments. In this case, in connected workflows, the same manipulation can influence code, messages, documents, tickets, cloud actions, or data transfer across trusted systems.
Prompt injection happens when untrusted content contains instructions that compete with the user’s real request. The danger rises when the assistant can retrieve private information, call tools, run commands, or make changes.
Every client remains isolated, access is attributable to a named operator, and the agency can deliver consistent evidence without revealing another client.
Context decision
Three questions before adding access
Can this operator or tool reach any repository, mailbox, drive, cache, token, or transcript belonging to another client?
Are credentials and AI sessions issued per client and person rather than shared across the agency?
Can the agency deliver useful proof to this client without including another client's names, paths, findings, or configuration?
Evidence goal: Create a separate client evidence record covering operator identity, workspace isolation, credentials, approved systems, review history, and delivery status.
A repeatable review
Four steps, no sensitive data required
- 1
Write down the exact MCP-Connected AI Assistants account, workspace, project, device, and connected service used in this workflow.
- 2
Inspect new or changed tools and run untrusted resources with no write, network, or secret authority.
- 3
Assign the decision and next review to the client service owner or agency security lead; do not leave the access boundary as an unwritten assumption.
- 4
Require explicit consent and treat server content as untrusted input to a separately enforced policy. Record the result without copying private content or raw credentials into the report.
Controls to apply
Reduce access before adding trust
Require explicit consent and treat server content as untrusted input to a separately enforced policy.
Separate trusted instructions from retrieved or user-supplied content.
Use tool allowlists, denied paths, network restrictions, and approval gates around consequential actions.
Log the source of instructions and stop when tool behavior changes unexpectedly.
Decision rule
Know when a formal baseline is justified
Simple text-only use still needs judgment, but the paid security case begins when untrusted content and meaningful tool authority coexist. That is the point to map the full action-to-asset path.
CapitalGuard is relevant when the workflow includes repositories, recurring private work, credentials, connected systems, commands, or evidence that must be shared with another person. It does not inspect this account from the page or guarantee that an incident cannot occur.
Primary references