Back to skill

Security audit

homebutler

Security checks for vulnerabilities and agentic risk

Overview

The verdict is benign with medium confidence: scanners found no AIG issues, while Skillspector flagged application management, deployment, and notification-test commands and concerns about acting without confirmation.

The verdict is benign, but confidence is medium and the Skillspector scan was partial because several referenced paths were missing. Before installing, consider whether you want a skill whose documented commands include server changes such as software installation, deployment, upgrades, and removal, and review any high-impact action before allowing it.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (5)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The supplied chunk exposes application lifecycle operations that change server state, beyond the declared monitoring and status-reporting purpose. The general statement about classifying write and destructive tools does not explicitly disclose application installation, removal, or purging. The listed commands also lack the promised classifications. This chunk documents commands; it does not establish that any were executed.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
87% confidence
Finding

Initialization, trust setup, and watching could support the declared monitoring purpose. However, serving, deployment, installation, and upgrade commands expose operational capabilities not adequately described by reporting server changes and status. Mentioning read/write/destructive classifications does not itself disclose these capabilities. The chunk also does not show those classifications. Exact side effects cannot be verified without the Homebutler implementation, but the command-level mismatch is evident.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
87% confidence
Finding

The supplied chunk performs a notification-test action, not one of the declared monitoring or management functions. Although notifications could support monitoring, this standalone test action has no visible connection to reporting server changes. The chunk does not reveal the notification destination or implementation.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 29)May include surrounding context.

md
Start `homebutler mcp` and call tools. Every tool is classed **read**, **write**
or **destructive**, and that classification is what lets an agent decide what it
may do without asking. It says what a call does to the machine, not what the
answer contains; what a read returns is under the classes, below. Shell
commands are for the handful of things no tool exposes, and for anything the
operator has to run themselves.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 68)May include surrounding context.

md
### What the classes mean for you

- **read** — changes nothing on the machine, so running one needs no
  confirmation. What comes back is another matter: hostnames, internal
  addresses, what is listening, what is running, usernames, host paths, log
  contents. `docker_top` names each process inside a container by the kernel's