Description-Behavior Mismatch
- Category
- MCP Tool Poisoning
- Confidence
- 99% confidence
- Finding
所提供代码没有读取财报或公告,也未执行条款提取、原句引用或来源标注。其实际功能是获取并打印外部 MCP 目录,且未体现该目录如何支持所声明的财报披露登记任务,因此属于实质性的目的与资源访问不匹配,而非可确认的辅助实现细节。
- Content
Security audit
Security checks for vulnerabilities and agentic risk
ClawScan rates this skill suspicious because scanners report a mismatch between its stated financial-disclosure purpose and some network activity, along with risks involving credentials and document uploads.
Consider avoiding installation until the MCP activity and its relationship to the stated task are clarified, the destination for credential-bearing requests is validated, and external document uploads are clearly disclosed and authorized. Scan coverage was partial: three of four components were scanned, and one static analysis was degraded, so the evidence is incomplete.
所提供代码没有读取财报或公告,也未执行条款提取、原句引用或来源标注。其实际功能是获取并打印外部 MCP 目录,且未体现该目录如何支持所声明的财报披露登记任务,因此属于实质性的目的与资源访问不匹配,而非可确认的辅助实现细节。
该代码块的实际用途是发现 MCP 服务路由和枚举远程工具,没有读取财报、摘录披露、生成登记表或建立原文引用的行为。所示内容也未表明这些远程服务与声明的财报披露整理任务有关,因此不能将其视为已明确关联的辅助实现。代码仅输出目录中的 routing,并未直接将其赋给 ROUTING_URL;后续请求依赖已有环境配置。无法据此认定凭据被恶意窃取,但外部服务发现和携带凭据的工具枚举属于声明中未体现的实质能力。
The anonymous catalog GET does not itself transmit documents or credentials, and service discovery is relevant to the skill. However, the surrounding workflow instructs sending CUE_API_KEY to a URL derived from the catalog without specifying destination validation, so a compromised or manipulated routing response could cause credential disclosure to an unauthorized endpoint.
发现先行,匿名取 live 目录,拿到 routing 后再连接发 tools/list,以现场返回为准,不硬编码工具名与工具数:
curl -sS https://cuecue.cn/api/mcp-catalog -o /tmp/mcp-catalog.json
python3 -c "import json;d=json.load(open('/tmp/mcp-catalog.json'));print(d.get('routing') or d)"
# 拿到 routing 后,带 token 发 tools/list:
curl -sS -X POST "$ROUTING_URL" -H "Authorization: Bearer $CUE_API_KEY" \
The workflow directs long attachments and scans to an external parsing service without explicitly warning that document contents leave the user's environment or requiring authorization. Published prospectuses present limited confidentiality risk, but user-supplied attachments could contain confidential annotations or unpublished financial information; the offline option does not itself establish informed consent.
Without declared permissions the skill's intent is opaque and cannot be validated.
The skill directs the agent to initiate external parsing when source text is inside an attachment, without an explicit disclosure warning or authorization check. Although many intended reports are public, user-provided PDFs or scans may contain confidential material that would consequently be transmitted outside the user's environment.
The checker rejects registered text whenever “递延” co-occurs with an expense-treatment phrase, even when that text faithfully reproduces the cited disclosure. This imposes an accounting conclusion contrary to the skill’s source-preservation purpose and also flags negated statements such as “递延利息不计入费用”; an agent trying to pass validation could consequently omit or alter accurate financial disclosures.