Security audit

Memory-Dream

Security checks across malware telemetry and agentic risk

Overview

This memory-cleanup skill is purpose-aligned, but it can change persistent memory files automatically while parts of its docs wrongly imply daily logs are read-only.

Review before installing or enabling cron. Run it manually first on a backed-up or version-controlled workspace, inspect diffs to MEMORY.md and memory/*.md, and prefer explicit phrases like "consolidate memory" over the broad "dream" trigger. This does not show exfiltration or destructive behavior, but it can quietly reshape persistent agent memory.

SkillSpector

By NVIDIA

Vulnerability Patterns

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
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access

Findings (5)

Intent-Code Divergence

Medium

Confidence: 94% confidence
Finding: The README makes a safety claim that daily memory files are read-only, but the documented workflow explicitly says old daily logs are modified to add consolidation markers. This inconsistency can mislead operators about what the skill changes, weakening informed consent, auditability, and backup expectations for files the user may consider immutable logs.

Intent-Code Divergence

Medium

Confidence: 95% confidence
Finding: The English notes state the skill is read-only on daily files, while earlier sections describe adding consolidation tags to old daily logs. Contradictory documentation about file mutation is dangerous because users may deploy the skill under the false assumption that source memory logs are never altered.

Vague Triggers

Medium

Confidence: 92% confidence
Finding: The trigger phrase "dream" is a common natural-language term and is likely to appear in ordinary conversation unrelated to memory consolidation. In an agent environment, this can cause accidental invocation of a state-modifying skill that reads memory files and edits MEMORY.md, creating unintended file changes and trust-boundary violations.

Vague Triggers

Medium

Confidence: 88% confidence
Finding: The automatic trigger conditions are loosely defined (e.g. Cron and heartbeat-based thresholds) without clear authorization or execution boundaries. This increases the chance of the skill running unexpectedly and modifying persistent memory based on routine system state rather than an explicit user-approved action.

Missing User Warnings

Medium

Confidence: 90% confidence
Finding: The skill text does not prominently disclose that it will modify MEMORY.md and append run logs to daily memory files, even though these are persistent state changes. Users may invoke the skill expecting analysis only, while it actually performs writes, which can surprise operators and enable prompt-triggered unauthorized modifications.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.