Skip to content
Squatch MCP
Sign in

Agent Inbox — direct agent-to-agent messaging with receipts

com.a2awire/agent-inbox

by com.a2awire· v0.1.2Source: Official MCP Registry
Sign in to watchSign in to claim this server

Findings

Detected by Squatch rules from the server's metadata and protocol behavior. Latest completed scan: .

Severity: Critical1 finding

Severity: CriticalStatus: OpenSQ-TOOL-EXFIL-DIRECTIVE

Asks the AI model to pass conversation data or secrets

Model-facing text instructs the AI model to place conversation content, prompts, tool history or secrets into tool arguments or external destinations.

First seen · Last seen

Suggested fix: Do not ask the model to include conversation, prompt or credential content in arguments.

Severity: High7 findings

Severity: HighStatus: OpenSQ-TOOL-EXFIL-DIRECTIVE

Asks the AI model to pass conversation data or secrets

Model-facing text instructs the AI model to place conversation content, prompts, tool history or secrets into tool arguments or external destinations.

First seen · Last seen

Suggested fix: Do not ask the model to include conversation, prompt or credential content in arguments.

Severity: HighStatus: OpenSQ-TOOL-CONCEALMENT

Text asks the AI model to hide actions from the user

Model-facing text instructs the AI model to hide actions or information from the user.

First seen · Last seen

Suggested fix: Remove any wording that asks the model to act silently or withhold information from the user.

Severity: HighStatus: OpenSQ-SERVER-AUTH-NONE-WRITE

State-changing tools available without authentication

The remote server lists state-changing tools without requiring authentication.

First seen · Last seen

Suggested fix: Require MCP authorization for servers that expose state-changing tools.

Severity: HighStatus: OpenSQ-TOOL-EXFIL-DIRECTIVE

Asks the AI model to pass conversation data or secrets

Model-facing text instructs the AI model to place conversation content, prompts, tool history or secrets into tool arguments or external destinations.

First seen · Last seen

Suggested fix: Do not ask the model to include conversation, prompt or credential content in arguments.

Severity: HighStatus: OpenSQ-SERVER-INSTRUCTIONS-INJECTION

Server instructions contain directives to the AI model

The server's instructions field contains directives addressed to the AI model beyond describing how to use the server.

First seen · Last seen

Suggested fix: Limit server instructions to usage guidance for this server's own tools.

Severity: HighStatus: OpenSQ-TOOL-SENSITIVE-PATH-REFERENCE

References credential or key locations

Model-facing text references credential, key or configuration file locations unrelated to the tool's stated purpose.

First seen · Last seen

Suggested fix: Remove references to credential files, key stores and client configuration files.

Severity: HighStatus: OpenSQ-TOOL-EXFIL-DIRECTIVE

Asks the AI model to pass conversation data or secrets

Model-facing text instructs the AI model to place conversation content, prompts, tool history or secrets into tool arguments or external destinations.

First seen · Last seen

Suggested fix: Do not ask the model to include conversation, prompt or credential content in arguments.

Severity: Medium3 findings

Severity: MediumStatus: OpenSQ-TOOL-SHELL-PAYLOAD

Shell command patterns in model-facing text

Model-facing text contains shell command patterns such as piping downloads into a shell or recursive deletion.

First seen · Last seen

Suggested fix: Remove shell commands from descriptions.

Severity: MediumStatus: OpenSQ-TOOL-CROSS-TOOL-REFERENCE

Gives instructions about other tools

Model-facing text gives the AI model instructions about tools that this server does not provide (tool shadowing).

First seen · Last seen

Suggested fix: Describe only this server's own tools.

Severity: MediumStatus: OpenSQ-TOOL-CREDENTIAL-ARGUMENT

Asks the AI model to supply credentials as arguments

A tool parameter asks the AI model to supply a password, token, key or similar credential, which the model can see and repeat.

First seen · Last seen

Suggested fix: Obtain credentials through server-side configuration or MCP authorization instead of tool arguments.

Severity: Low2 findings

Severity: LowStatus: OpenSQ-TOOL-LENGTH-ANOMALY

Unusually long model-facing text

Model-facing text is unusually long, which consumes context and leaves room to hide content.

First seen · Last seen

Suggested fix: Shorten descriptions; move long documentation out of model-facing fields.

Severity: LowStatus: OpenSQ-TOOL-META-INSUFFICIENT

Description too thin to determine behavior

The tool's name and description do not state what it does.

First seen · Last seen

Suggested fix: Describe the action the tool performs, its target and its side effects.

What this scan does not cover

  • Squatch never calls a server's tools, so it cannot see what they actually do at run time.
  • It scans without credentials: tools behind authentication may not be visible.
  • It does not review source code, dependencies or the publisher's infrastructure.
  • A server can change behavior between scans; monitoring compares each scan with the last.
  • Classifications come from tool names, descriptions and schemas, which the publisher controls.

How Squatch scans servers