Security audit

Coc Helper

Security checks across malware telemetry and agentic risk

Overview

This is a local Call of Cthulhu tabletop helper that stores game state in its own JSON file and does not show evidence of hidden network, credential, or system-level behavior.

Install only if you are comfortable with character and session data being saved locally in the skill's session.json. Avoid putting private real-world information in investigator names or notes, and be aware that one SAN threshold helper message appears inaccurate compared with the main documented rules.

SkillSpector

By NVIDIA

Vulnerability Patterns

Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
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
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access

Findings (4)

Intent-Code Divergence

Medium

Confidence: 97% confidence
Finding: The helper text in `sanThreshold()` states that dropping below the SAN/5 threshold causes indefinite insanity, but the implemented `sanCheck()` logic says indefinite insanity is triggered by a single loss above a configured threshold or by caller-tracked cumulative daily loss. This inconsistency can mislead users and downstream tooling into applying the wrong game-state transitions, causing persistent state corruption or incorrect adjudication in a rules assistant that advertises investigator state persistence and JSON chaining.

Missing User Warnings

Low

Confidence: 92% confidence
Finding: The proposal explicitly describes persistent investigator/session state written to JSON, but the user-facing description does not clearly warn that game data will be stored on disk. This can surprise users and lead to unintended retention of names, notes, or campaign state on shared machines, though the data appears local-only and not exfiltrated.

Vague Triggers

Medium

Confidence: 88% confidence
Finding: The proposal defines activation examples in natural language that are broad enough to overlap with ordinary conversation, which can cause the agent to invoke the skill when the user did not clearly intend it. In this skill's context, unintended invocation is moderately risky because the tool is stateful and can perform local writes, so a casual mention of dice, NPCs, or checks could trigger persistence or game-state changes without explicit user consent.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The document states that investigator and session data are persisted to a local JSON file, but it does not clearly warn users at the point of use that their inputs and state will be written to disk. In a stateful assistant, this creates a privacy and consent issue because users may disclose names, notes, or campaign details assuming an ephemeral interaction, while the skill silently stores them locally for later reuse.

VirusTotal

65/65 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.