Back to skill

Security audit

yu

Security checks for vulnerabilities and agentic risk

Overview

The skill does not run code or access the system, but it can add unverified academic or medical analysis while presenting itself as language optimization.

Review this skill carefully before installing if you need faithful copyediting. It is best suited for aggressive Chinese academic rewriting, not conservative polishing. Any added evidence language, mechanisms, clinical implications, theoretical insights, or authorial stance should be checked against the original sources before use in a manuscript.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

other

Warning
Location
SKILL.md:14
Finding

Academic Content Integrity Risk from Unsupported Substantive Additions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 14–93
Vulnerability Type: other: Academic content fabrication and integrity manipulation
Risk Level: Medium

Complete Relevant Source Snippet

The following is an English translation of the relevant source directives:

markdown
This skill's language optimization must not change the core viewpoints, data, or conclusions of the original text.

On the basis of round-trip translation, identify all data, mechanisms, and concluding statements in the original text, and require the addition of evidence-aware or more precise wording; meanwhile, construct a clear “problem–mechanism–limitation–need–outlook” logical chain.

Add brief, analytical connective statements between paragraphs or within sentences to simulate the thought process of a real researcher.

Add words or short phrases that express the author's position and demonstrate the author's interpretation and evaluation of the material.

Conduct a final in-depth review of the full text: require deeper mechanism descriptions, providing specific neural or behavioral pathways rather than generalized terminology; strengthen the researcher's personal academic position, avoid formulaic traces, ensure a complete and natural logical chain, and add original critical analysis.

Delete or reconstruct sections containing accumulated concepts; add targeted clinical or theoretical insights; integrate hedging naturally rather than mechanically.

Technical Analysis

The skill initially states that optimization must preserve the source's viewpoints, data, and conclusions. Later instructions, however, require the model to add authorial positions, original critical analysis, specific neural or behavioral mechanisms, and targeted clinical or theoretical insights.

These are substantive transformations rather than purely linguistic edits. No corresponding requirement instructs the model to verify the additions against u ...[truncated 2141 chars]

Remediation
View remediation

Remediation Suggestions

  1. Restrict the default workflow to wording, grammar, sentence structure, organization, and clarity.
  2. Prohibit inventing or adding data, citations, mechanisms, clinical implications, theoretical insights, conclusions, or authorial positions not present in the source.
  3. Replace mandatory substantive additions with optional recommendations shown separately from the edited text.
  4. Require explicit user approval before incorporating any new interpretation or mechanism.
  5. Require a verifiable source for every evidence-dependent addition; if no source is available, label the statement clearly as an unverified proposal.
  6. Preserve citations exactly and prohibit creating placeholder, inferred, or fabricated references.
  7. Add a final semantic-diff check confirming that all factual claims, numerical data, conclusions, and causal relationships remain faithful to the input.
  8. For medical content, require domain-expert review before adding clinical implications or specific neural and behavioral pathways.
  9. Clearly distinguish among direct edits, optional analytical suggestions, and statements requiring external verification.
Vulnerability Patterns
  • 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
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (3)

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest description and overview are entirely framed around Chinese-language academic optimization and explicitly describe converting output back into high-quality Chinese. There is no opt-in or alternative for users who may want another language, which constitutes a language/locale policy issue under the stated rules.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill is presented as a language optimizer, but its workflow directs the model to construct logic chains, strengthen evidence awareness, deepen mechanisms, and inject researcher viewpoint. This creates a hidden scope expansion where users may believe they are receiving harmless copy-editing while the model is actually generating new analytical content, increasing the risk of factual drift and academic integrity issues.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill states that optimization will not change core claims, data, or conclusions, yet later instructs the agent to add evidentiary framing, researcher stance, speculative hedging, and analytical bridging text. That mismatch can silently alter meaning, introduce unsupported implications, and misrepresent the provenance of academic content even if no external code execution is involved.

Content

No source excerpt is available for this finding.