Description-Behavior Mismatch
- Category
- MCP Tool Poisoning
- Confidence
- 99% confidence
- Finding
所提供代码仅获取并展示外部 MCP 目录,没有读取财报或公告、提取会计政策变更信息或生成带原句引用的登记表。该网络访问也未体现为已声明任务的必要支持步骤,因此实际行为与声明用途存在实质不符。代码未显示上传财报或其他本地数据的行为。
- Content
Security audit
Security checks for vulnerabilities and agentic risk
The skill is rated suspicious because scanners report behavior that does not match its accounting-policy disclosure purpose and could weaken the reliability of its checks; some analysis was incomplete.
Consider avoiding installation until the credential destination, external document handling, and validator behavior are clarified. The scan was partial: 3 of 4 components were fully inspected, and analysis of the fixture runner was limited, so the results do not establish that unreviewed behavior is safe.
所提供代码仅获取并展示外部 MCP 目录,没有读取财报或公告、提取会计政策变更信息或生成带原句引用的登记表。该网络访问也未体现为已声明任务的必要支持步骤,因此实际行为与声明用途存在实质不符。代码未显示上传财报或其他本地数据的行为。
所给代码执行的是远程 MCP 服务发现和工具枚举,没有读取财报、提取会计变更披露或建立原文可追溯登记表。代码也未展示该服务与声明的会计披露整理任务存在关联,因此这些网络访问和凭证使用不能仅视为已声明功能的支持性实现细节。此外,代码仅打印 routing,并未将其赋给 ROUTING_URL;后续请求实际使用已有环境变量指定的地址。
登记稿校验属于所声明披露整理任务的合理辅助功能,本身不构成实质偏离。但 main() 无条件调用 mark_first_run(),在实际处理文档前向任务输入之外的用户主目录写入每日使用标记。这一持久化使用跟踪行为未在描述或权限中声明,也不是实现登记表或校验所必需的。该行为是本地记录,不应误称为远程遥测或数据外传。
Referenced artifact was not completely inspected
| `scripts/fixtures/run_fixtures.sh` | 离线样例 runner:自测→好样→坏样三档,注入必先断言落地 |
Anonymous HTTPS catalog retrieval is not itself exfiltration, and authenticated tool discovery is relevant to the workflow. However, the instructions direct the agent to send CUE_API_KEY to a routing URL obtained from the catalog without specifying destination validation; a compromised catalog or incorrectly supplied routing value could expose the credential to an unintended host.
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 recommends sending attachments to an external parsing service without requiring a confidentiality check or informed user approval. Published financial reports present limited privacy risk, but user-supplied drafts, screenshots, or annotated documents could contain nonpublic financial information.
Without declared permissions the skill's intent is opaque and cannot be validated.
The skill directs external parsing when evidence resides in attachments but does not require a privacy check or approval before uploading user-supplied documents. Most published reports are public, which reduces typical exposure, but confidential drafts or annotated attachments could be disclosed outside the user's intended trust boundary.
An entry containing 〔自算〕 bypasses both the quotation requirement and the amount/source consistency checks, allowing fabricated figures to receive a clean result if the remaining document checks pass. Because this skill specifically promises disclosed, source-backed figures rather than independently calculated amounts, this attacker-controlled exemption undermines its central integrity guarantee; the remaining keyword checks do not reliably prevent the bypass.
The checker creates a persistent, dated skill-usage marker under ~/.cue-usage without obtaining consent, including during --demo and --selftest runs. This is unnecessary for document validation and exposes limited local usage metadata; however, the marker is empty and the code neither transmits it nor records financial document contents.