Security audit

Self Improving Agent

Security checks across malware telemetry and agentic risk

Overview

This skill is mostly about agent memory, but it can automatically persist conversation details, raw errors, and project guidance without clear redaction or per-entry user consent.

Install only if you want an agent to keep project memory on disk. Treat .learnings and promoted instruction files as potentially sensitive: review entries before committing, redact tokens and personal or customer data, and prefer explicit prompts like 'log this learning' over automatic capture.

SkillSpector

By NVIDIA

Vulnerability Patterns

Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access

Findings (5)

Vague Triggers

High

Confidence: 92% confidence
Finding: The trigger keywords are extremely broad and overlap with ordinary conversation, making accidental activation likely. In a skill that can write persistent files and recommend hook/execution workflows, over-broad triggers increase the chance of unintended logging, instruction injection into project memory, or surprise file modifications during unrelated tasks.

Ssd 3

Medium

Confidence: 96% confidence
Finding: The learning-entry format encourages storing full context, detailed descriptions, related files, and conversation-derived metadata in persistent local files without any minimization guidance. This can capture secrets, personal data, proprietary code details, or sensitive business context and retain them long-term, increasing exposure through commits, backups, sharing, or later model access.

Ssd 3

Medium

Confidence: 98% confidence
Finding: The error logging template explicitly asks for raw error output plus inputs and parameters, which frequently contain API keys, auth headers, file paths, customer data, prompts, or other confidential material. Persisting this verbatim creates a direct data-leak risk, especially if the files are committed or read by other tools and agents later.

Ssd 3

Low

Confidence: 86% confidence
Finding: The feature-request template tells the agent to preserve the user's requested capability and context, which can needlessly retain sensitive product plans, internal workflows, or personal details. While lower risk than raw error dumps, it still creates persistent storage of user-supplied information without minimization or consent.

Ssd 3

Medium

Confidence: 94% confidence
Finding: The automatic detection triggers create a standing instruction to log user corrections, missing information, and errors whenever they are noticed, effectively turning ordinary conversation into persistent retention. This is dangerous because it normalizes saving user-provided content without an explicit consent gate, and the triggers are broad enough to capture sensitive facts during normal troubleshooting or planning.

VirusTotal

VirusTotal findings are pending for this skill version.

View on VirusTotal

Static analysis

No suspicious patterns detected.