Security audit

Neosoul Decision Agent

Security checks across malware telemetry and agentic risk

Overview

This decision helper is not malicious, but it should be reviewed because it automatically builds and reuses local memory files that may contain sensitive personal or business context.

Install only if you are comfortable with the agent keeping a local decision profile under ~/decision-making. Review or delete those files regularly, avoid recording sensitive personal, legal, health, or business strategy details unless intentional, and ask the agent to confirm before saving new preferences or decision records.

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 (9)

Intent-Code Divergence

Medium

Confidence: 93% confidence
Finding: The skill claims it will never modify its own SKILL.md, yet it exposes the exec tool and includes shell-based file creation and update workflows. That mismatch creates a trust-boundary problem: the model may be induced to run arbitrary file-modifying commands, including commands that alter the skill definition itself or other local files, despite the documented safety claim.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The skill persistently stores user preferences and decision history to local files, but the upfront description does not clearly warn users that their conversation-derived preferences and retrospective records will be retained. This is dangerous because users may disclose sensitive personal or business context under the assumption of ephemeral processing, leading to privacy leakage on the local machine or unintended long-term profiling.

Vague Triggers

Medium

Confidence: 90% confidence
Finding: The activation language is broad enough that ordinary discussion of options, hesitation, or tradeoffs could trigger memory updates or proactive analysis without clear user intent. In a skill that persistently writes inferred preferences and decision history, over-broad triggering increases the chance of collecting and storing sensitive data from casual conversation without meaningful consent.

Vague Triggers

Medium

Confidence: 91% confidence
Finding: The proactive detection criteria are very general and lack clear exclusions, so the skill may initiate analysis and reference prior stored preferences in situations where the user did not request persistent profiling or intervention. Given the memory system, this broad scope raises privacy and consent risks and can cause the agent to over-harvest conversational data into long-lived records.

Ssd 3

Medium

Confidence: 96% confidence
Finding: The skill is designed to persist and reuse user preference and history data across future interactions, effectively creating a local behavioral profile. This is dangerous because decision preferences, regret history, and bias annotations can be highly sensitive and may reveal personal, professional, or psychological information if accessed by other local users, malware, backups, or later prompts.

Ssd 3

Medium

Confidence: 95% confidence
Finding: The HOT/WARM/RECORD/COLD design instructs the agent to always load and retain persistent user preference data and key rules, which normalizes long-term profiling as a default behavior. Always-loading memory increases exposure of sensitive historical data to every future interaction and raises the chance of accidental disclosure, over-personalization, or prompt-induced leakage from prior sessions.

Ssd 3

Medium

Confidence: 97% confidence
Finding: The skill explicitly directs the agent to infer signals from conversation and write them into persistent memory and domain files. Even if framed as user preference learning, this creates a covert data collection risk because the system turns natural language statements into durable profile entries that may outlive the original context and be reused in later decisions.

Ssd 3

Medium

Confidence: 96% confidence
Finding: The retrospective records include decision context, outcomes, lessons, and detected cognitive biases, which can capture highly sensitive personal, commercial, or mental-state information. Persisting this level of reflective detail creates a richer dossier than ordinary chat logs and can materially harm users if exposed, especially in workplace, health, relationship, or strategic business contexts.

Ssd 3

Medium

Confidence: 95% confidence
Finding: The usage flow operationalizes automatic memory updates and dated decision-history file creation from ongoing user interactions, making persistent collection part of the default workflow. This is dangerous because it systematizes storage of potentially sensitive conversational material without strong gating, increasing the likelihood of privacy violations and unintended long-term retention.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.