mayubench-en

Security checks across malware telemetry and agentic risk

Overview

This skill is a Markdown-only AI behavior benchmark, with risky phrases and sensitive scenarios used as test cases rather than hidden actions.

Safe to install as a reference benchmark. Use it intentionally for evaluations, keep its adversarial prompts separate from live user workflows, and treat the hard-coded crisis-resource example as benchmark text rather than production safety guidance.

SkillSpector

By NVIDIA

Vulnerability Patterns

Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (2)

Natural-Language Policy Violations

Medium

Confidence: 88% confidence
Finding: The benchmark hard-codes a specific crisis hotline number and assumes a city-based support path without checking the user's location or whether the resource is appropriate. In a real deployment, this can misdirect a person in acute mental-health crisis to unavailable or irrelevant resources, delaying access to effective help.

Vague Triggers

Medium

Confidence: 82% confidence
Finding: The line "behavior questions will automatically trigger MayuBench evaluation" describes broad, implicit activation without defining what counts as a trigger, when evaluation starts, or whether the user must explicitly opt in. In an agent-skill environment, ambiguous auto-triggering can cause the skill to intercept unrelated prompts, override normal agent behavior, or run evaluations unexpectedly, which creates prompt-routing and scope-control risk even though the skill appears benchmark-focused rather than overtly malicious.

VirusTotal

65/65 vendors flagged this skill as clean.

View on VirusTotal