Security audit

交付物发前体检

Security checks across malware telemetry and agentic risk

Overview

This skill is a coherent pre-send document checker that inspects business materials for send-risk issues, with no evidence of hidden exfiltration, destructive behavior, or unsafe persistence.

Use this skill only for materials you are comfortable having fully inspected, including hidden slides, sheets, comments, notes, and metadata-like review traces. Prefer passing only the specific files or attachments intended for the send-readiness check, and avoid running raw extraction helpers in environments where stdout is logged or shared.

SkillSpector

By NVIDIA

Vulnerability Patterns

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

Findings (5)

Vague Triggers

Medium

Confidence: 91% confidence
Finding: The adapter defines several loose activation phrases such as '这份能发吗', '帮我做出街体检', and '检查一下', which can overlap with ordinary workplace requests and may cause the skill to trigger when the user did not explicitly intend a pre-send gate. Because this skill changes behavior based on inferred send-readiness context and may generate authoritative release judgments or sending language, accidental invocation can lead to over-collection of context, inappropriate review framing, or misleading go/no-go recommendations in the wrong workflow.

Missing User Warnings

Medium

Confidence: 88% confidence
Finding: The skill explicitly instructs inspection of hidden worksheets, comments, revision traces, and export artifacts, which can expose sensitive or unintended-to-share information beyond the obvious visible document body. In a pre-send review skill this access is functionally relevant, but without a clear user-facing consent boundary or minimization rule, it creates a privacy/data-exposure risk by encouraging broader document inspection than a user may expect.

Missing User Warnings

Medium

Confidence: 85% confidence
Finding: The guidance to inspect attachments, compressed packages, and same-directory materials expands access scope from the primary file to surrounding files that may contain unrelated or more sensitive content. That broader context can be useful for cross-attachment consistency checks, but without explicit user warning and scope confirmation, it risks over-collection and accidental review of files the user did not intend to expose.

Natural-Language Policy Violations

Medium

Confidence: 88% confidence
Finding: The contract requires Chinese business labels in the report body and unconditionally constrains output language, which can override or conflict with the user's language preference. In a pre-send review skill, this can cause usability, accessibility, or policy-compliance issues if the recipient context or operator expects another language, and it may lead to incorrect downstream handling or missed review signals.

Missing User Warnings

Medium

Confidence: 88% confidence
Finding: The script extracts full text from PPT, Word, Excel, PDF, and pasted-text inputs and emits the aggregated contents directly to stdout as JSON. In a pre-send review skill, those files are likely to contain confidential business data, comments, notes, hidden slides, and other sensitive material, so writing everything to stdout can unintentionally expose contents through logs, parent-process capture, shell history workflows, or downstream tooling not intended to store raw documents.

VirusTotal

64/64 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

No suspicious patterns detected.