Security audit

Assess Me

Security checks across malware telemetry and agentic risk

Overview

This reasoning-audit skill is mostly transparent, but it asks agents to write internal thinking to disk and to modify the skill itself, which needs human review before installation.

Review this skill before installing. It does not appear malicious and VirusTotal/static scan signals are clean, but users should be comfortable with agents writing audit state to a local temp file and with the skill's self-optimization language. Prefer a version that keeps the audit in memory or writes only sanitized summaries, deletes temporary files, and removes automatic skill-manager repair or optimization steps.

SkillSpector

By NVIDIA

Vulnerability Patterns

Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
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

Findings (5)

Context-Inappropriate Capability

High

Confidence: 97% confidence
Finding: The skill explicitly instructs loading `skill-manager` to repair or optimize itself based on verification results, which expands a reasoning-audit skill into self-modification behavior. Self-modifying or self-extending behavior is dangerous because it can change future execution paths, import unrelated capabilities, and create a pivot from benign introspection into uncontrolled maintenance actions without clear user approval.

Description-Behavior Mismatch

Medium

Confidence: 94% confidence
Finding: The Verification section broadens the documented purpose from auditing current reasoning into maintaining and improving the skill itself. This scope creep is risky because it normalizes actions outside the declared purpose, making it easier for the skill to trigger capability expansion or modification under the guise of routine verification.

Missing User Warnings

Medium

Confidence: 96% confidence
Finding: The skill requires writing the agent's current thinking state to `/tmp/assess-me.md` and reading it back, creating persistent local storage of potentially sensitive internal reasoning. This is dangerous because internal reasoning may contain secrets, user data, security assumptions, or system context that should not be written to disk, especially without minimization, consent, or cleanup guarantees.

Vague Triggers

Medium

Confidence: 90% confidence
Finding: The trigger rules activate on common phrases like 'review' and 'verify' or general user doubt, making invocation too easy and potentially frequent. Overbroad activation is risky here because it can unnecessarily invoke the file-writing introspection workflow and expand the skill's reach into normal conversations where such behavior was not intended.

Vague Triggers

Medium

Confidence: 91% confidence
Finding: The trigger table includes broad and subjective conditions such as ambiguous requirements, unverified assumptions, or vague integrity checks, which gives the skill large discretion to activate itself. In context, that increased autonomy makes the skill more dangerous because it can initiate introspective file operations and subsequent workflow changes without a clear, narrow boundary.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.