Security audit

Craw & Core — Lobster Dungeon

Security checks for vulnerabilities and agentic risk

Overview

This is mostly a coherent networked game skill, but it automatically creates a recurring daily reminder without a clear opt-in or disable path.

Install only if you are comfortable with a networked game sending your platform ID and game activity to craw-core.com and creating a daily reminder after character creation. Review or remove the reminder job named daily-lobster-reminder if you do not want ongoing notifications.

SkillSpector

Vulnerability Patterns

Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
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

Findings (5)

Context-Inappropriate Capability

Medium

Confidence: 94% confidence
Finding: The skill instructs the agent to create a persistent daily cron reminder automatically after walker creation, without any explicit user consent or clear disclosure. This expands the skill from a passive game interaction into ongoing background engagement, which can create unwanted persistence, notification fatigue, and privacy/trust concerns if the user did not opt in.

Description-Behavior Mismatch

Medium

Confidence: 88% confidence
Finding: The manifest presents the skill as a game, but the documentation includes subscription-status checks and activation-code redemption flows, which are payment-adjacent capabilities outside the minimal declared purpose. This broadens the trust boundary and could lead users to expose purchase or account-related data in a context where they may not expect commerce handling.

Missing User Warnings

Medium

Confidence: 96% confidence
Finding: Mandating automatic daily reminders without warning or opt-in is a clear consent and user-expectation problem. Persistent proactive messaging can be abused for engagement farming or nuisance behavior, especially because it continues beyond the immediate user request flow.

Natural-Language Policy Violations

Medium

Confidence: 79% confidence
Finding: Forcing the first message to always be in English ignores the user's likely language preference and removes an expected degree of user control. While not a severe exploit, it can degrade usability, create confusion, and violate platform expectations around adaptive language behavior.

Ssd 1

Medium

Confidence: 98% confidence
Finding: The skill includes semantic instructions telling the agent not to summarize or analyze the file and to immediately adopt the role of game master. This is a prompt-injection pattern designed to override the host agent's normal control flow and can suppress safety review, alter intended behavior, or trigger unauthorized actions such as onboarding and API usage.

Static analysis

No suspicious patterns detected.