Security audit

主观题评分标准(初高中语英作文+语文阅读理解)

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed classroom grading helper with local rubric files and no executable code, but users should supervise student-data handling and saved rubric changes.

Install only if you are comfortable using it for student work and local rubric files. Redact student identifiers when possible, confirm the rubric before grading, and do not let conversational feedback permanently modify criteria/*.json without reviewing the exact change and keeping a backup.

SkillSpector

Vulnerability Patterns

Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code

Findings (7)

Intent-Code Divergence

Medium

Confidence: 97% confidence
Finding: The file declares a full score of 60, but later defines score bands such as “二类上（59-62分）”, which exceeds the system’s own maximum. In an automated grading skill, this inconsistency can cause invalid outputs, broken normalization logic, policy bypasses, or downstream errors when consumers assume all scores are bounded by the declared maximum.

Intent-Code Divergence

Medium

Confidence: 95% confidence
Finding: The file presents two conflicting scoring-tier schemes for the same high school English writing standard: one section describes six levels, while the later 'official' rules state five tiers plus zero. In a grading skill, contradictory normative rules can directly cause inconsistent or unfair scoring, audit disputes, and unreliable automated evaluation outcomes.

Missing User Warnings

Medium

Confidence: 89% confidence
Finding: The README explicitly states that after grading, a teacher's instruction will automatically append new rules to files under criteria/*.json. Modifying local configuration/data based on natural-language input without an explicit confirmation step or warning is dangerous because a user may not realize persistent state is being changed, and malicious or mistaken input could silently corrupt future grading behavior.

Missing User Warnings

Medium

Confidence: 83% confidence
Finding: The skill says users can directly send images or text and that the AI will automatically recognize and process them, but it does not warn that student submissions may contain personal or sensitive educational data. In a grading context, this increases privacy risk because names, handwriting, school details, and performance data may be processed or retained without informed consent or minimization.

Missing User Warnings

Medium

Confidence: 95% confidence
Finding: The skill states that newly added grading criteria files will be 'automatically read' and later describes updating local criteria data as part of the workflow, but it does not clearly require explicit user confirmation before persistent file modification. This creates a real risk of unauthorized or unintended local state changes, configuration poisoning, or persistence of malformed/adversarial criteria that affect future grading runs.

Missing User Warnings

Medium

Confidence: 97% confidence
Finding: The '进化机制' explicitly says teacher feedback will automatically append data to criteria files and make it effective on future runs, again without an explicit safety warning, approval checkpoint, or integrity controls. In context, this makes the skill more dangerous because grading policy becomes persistent mutable state controlled through conversational input, enabling accidental corruption or deliberate prompt-based policy tampering across sessions.

Natural-Language Policy Violations

Medium

Confidence: 88% confidence
Finding: The criteria explicitly allow score deductions for traditional characters and uncommon online language without any user-selected locale, exam-board requirement, or accessibility justification. In a grading system, this embeds a hard-coded linguistic bias that can systematically and unfairly penalize valid regional or stylistic language forms, producing discriminatory outcomes at scale.

Static analysis

No suspicious patterns detected.