Back to skill

Security audit

事实核查

Security checks for vulnerabilities and agentic risk

Overview

The skill is marked suspicious because the scan reports an unpinned external runner that is executed with the agent’s permissions, alongside a mismatch between its fact-checking description and some reported behavior.

Consider avoiding installation until the external runner can be pinned to a reviewed version and the API key setup avoids sharing secrets in chat. The scan was partial: one local path reference could not be resolved, so the evidence is incomplete.

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
Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (8)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Document conversion could support delivery of a fact-checking report, but it is the entirety of the supplied behavior: the code assumes a Markdown report already exists and only changes its format. Thus, the supplied chunk does not implement the declared primary purpose. No network access, unrelated triggers, or other resource access is shown.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

Installing a runner is explicitly documented and supports the fact-checking workflow, so the description mismatch alone is not suspicious. However, the installation fetches or updates an unpinned repository and subsequently executes its Python code with the agent's permissions; a compromised upstream or mirror could thereby access local files and the Cue credential.

Content

No source excerpt is available for this finding.

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · SKILL.md (reported line 227)May include surrounding context.

md
CUE_KEY=$(python3 -c "import json;print(json.load(open('$HOME/.cue/config.json'))['api_key'])" 2>/dev/null || echo "$CUE_API_KEY")
echo "=== 1/4 API Key ===" && [ -n "$CUE_KEY" ] && echo "已配置" || echo "未配置!"
echo "=== 2/4 Cue 服务 ===" && curl -sS --max-time 10 "https://cuecue.cn/api/health" -H "Authorization: Bearer $CUE_KEY"
echo "=== 3/4 搭子 ===" && curl -sS --max-time 10 "https://cuecue.cn/api/playbook" -H "Authorization: Bearer $CUE_KEY" | python3 -c "import sys,json;scenes=json.load(sys.stdin).get('data',{}).get('scenes',[]);buddy=[b for s in scenes if s.get('secondary_category')=='深度核查' for b in s.get('buddies',[]) if b.get('title')=='事实核查'];print(f'可用:{len(buddy)}个') if buddy else print('暂不可用')"
echo "=== 4/4 剩余积分 ===" && python3 -c "
import json, urllib.request, os
p = os.path.expanduser('~/.cue/config.json')

The setup workflow requests an API key in chat without warning about credential exposure.

Medium
Category
Output Handling
Confidence
98% confidence
Finding

The setup repeatedly instructs users to paste a live API key into the conversation, exposing it to chat retention, logs, and any systems able to access that conversation. Authentication is necessary for Cue, but sending the secret through the assistant is avoidable and could enable unauthorized API use or credit consumption.

Content

No source excerpt is available for this finding.

The mandatory low-credit dialog forces Chinese output without a language choice.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The skill requires the Chinese question and options to be used verbatim and prohibits changing them. This mandates Chinese user-facing output regardless of the user's language, without offering language opt-in or documenting a locale-specific justification.

Content

No source excerpt is available for this finding.

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · SKILL.md (reported line 202)May include surrounding context.

md
| 目标格式 | 依赖 | macOS | Ubuntu |
|----------|------|-------|--------|
| Word (.docx) | pandoc | `brew install pandoc` | `sudo apt install pandoc` |
| PDF (.pdf) | pandoc + LaTeX | `brew install --cask basictex` | `sudo apt install texlive-xetex` |

---

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · SKILL.md (reported line 203)May include surrounding context.

md
| 目标格式 | 依赖 | macOS | Ubuntu |
|----------|------|-------|--------|
| Word (.docx) | pandoc | `brew install pandoc` | `sudo apt install pandoc` |
| PDF (.pdf) | pandoc + LaTeX | `brew install --cask basictex` | `sudo apt install texlive-xetex` |

---

The fact-checking workflow mandates referral and purchase promotion after successful reports, extending beyond the manifest's stated purpose.

Low
Category
Excessive Agency
Confidence
93% confidence
Finding

The skill makes referral and purchase promotion mandatory and prohibits unrelated questions in the same response, allowing vendor marketing instructions to redirect the assistant beyond the user's fact-checking request. This is a low-impact workflow-integrity issue rather than evidence of payment fraud: the promotion is conditional, targets the stated provider, and includes a decline option.

Content

No source excerpt is available for this finding.