Security audit

Human-Like Memory

Security checks across malware telemetry and agentic risk

Overview

This is a disclosed remote long-term memory skill, with privacy risks users should manage but no artifact-backed evidence of hidden or malicious behavior.

Install only if you want a remote long-term memory service. Avoid saving secrets, tokens, private command output, or sensitive tool results; review the configured endpoint and privacy policy; and consider disabling auto-save or HUMAN_LIKE_MEM_CAPTURE_TOOL_CALLS unless you explicitly need them.

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 (5)

Description-Behavior Mismatch

Medium

Confidence: 95% confidence
Finding: The code explicitly collects assistant tool calls and tool-result content into persistent context blocks when captureToolCalls is enabled, and that setting defaults to true. Tool arguments and results often contain secrets, file contents, command output, or internal system data, so persisting them materially expands data collection beyond the stated memory purpose and can create a sensitive-data retention channel.

Missing User Warnings

Medium

Confidence: 94% confidence
Finding: The save path sends conversation messages to a remote service endpoint with API credentials, but the CLI provides no explicit runtime consent, warning, or data-minimization step before transmission. Because this skill is designed to handle long-term memory and can include assistant/tool context, users may unknowingly disclose sensitive personal or operational data to a third party.

Missing User Warnings

Medium

Confidence: 89% confidence
Finding: Recall queries, user identifiers, agent identifiers, and scenario metadata are sent to a remote memory API without clear user notification at the point of use. Search queries can themselves contain highly sensitive information, so silent remote transmission creates a privacy and data-governance risk even if transport is HTTPS.

Vague Triggers

Medium

Confidence: 91% confidence
Finding: The trigger list is broad enough to match ordinary user phrases like 'remember what we discussed' or 'continue previous work', which can invoke persistent memory behavior unexpectedly. In a memory skill, accidental invocation is risky because it may cause sensitive prior-session data to be searched or saved without clear user intent.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The manifest explicitly advertises persistent long-term memory across sessions but provides no user-facing warning about retention, privacy, or what categories of data may be stored. Because this skill is designed to preserve facts, preferences, and decisions, the lack of disclosure increases the chance that personal or sensitive information is retained without informed consent.

VirusTotal

67/67 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.