Back to skill

Security audit

研发支出资本化·逐条挂源

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked this skill suspicious because scanners report undisclosed external service discovery, credential and document handling concerns, and local usage tracking that do not clearly fit its stated financial-disclosure purpose.

Consider the skill's suspicious verdict before installing. Review whether you are comfortable with its reported third-party document parsing, catalog-based MCP requests that carry an API key, and local usage marker. The AIG scan found no issues, but Skillspector was partial and reported unresolved references, so the evidence is incomplete; the findings do not by themselves establish malicious intent.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (23)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied commands perform service discovery and tool enumeration, not extraction or analysis of R&D capitalization disclosures. No report processing, quotation capture, citation tracking, or financial reconciliation appears in the chunk. Although tool discovery could support a larger workflow, this chunk does not establish any connection between these external services and the declared purpose, making the observed resource access and functionality materially inconsistent with the description.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The document-validation checks support the declared evidence-ledger purpose. However, main() calls mark_first_run() after argument parsing, before validation, demo output, self-tests, or checking whether an input path was supplied. That function creates a directory and dated marker outside the input document's location, suppressing any errors. This usage-recording side effect is unrelated to the declared financial-disclosure task and is not disclosed in the description. No network transmission is present in the supplied code.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

所示代码执行的是远程 MCP 服务发现与工具枚举,没有读取或分析研发披露、提取资本化金额与条件、计算比例或生成可追溯台账。代码也未展示这些网络请求与所声明财务分析任务之间的联系,因此不能将其认定为支持该任务的普通实现细节。未见代码上传财务文档或调用枚举出的工具,但已展示的外部资源访问和认证工具发现能力与声明用途存在实质不符。

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied chunk performs service discovery and tool enumeration, not extraction or verification of R&D disclosures. Although tool discovery could support a larger workflow, this code establishes no connection between the discovered services and the declared task. It neither processes report text nor generates source-linked accounting records, and its external service-discovery behavior is not represented in the declared purpose. The bearer token is sent to the configured endpoint; the snippet alone does not establish credential theft or other malicious intent.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

所给代码仅执行 MCP 服务与工具发现,没有读取年报、提取研发支出数据、核对资本化条件或生成带原句来源的台账。工具发现本身可能是某些系统的辅助步骤,但该片段未显示这些服务或工具与声明的财务披露分析任务存在联系;其明确实现的是未声明的外部服务探测和带凭据的工具枚举,因此与声明用途不符。代码未显示凭据被发送到非预期地址,也未显示财务数据外传。

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The CLI, fixture, and validation tests are supporting implementation details consistent with the declared purpose. However, the suite also explicitly requires a local usage-beacon side effect unrelated to extracting or validating financial disclosures. That behavior is not disclosed in the description or permissions. The beacon implementation is in the invoked checker, which is not supplied; this chunk establishes and tests the expected behavior rather than directly implementing it. No network transmission, credential access, or remote telemetry is shown.

Content

No source excerpt is available for this finding.

Env Variable Harvesting

High
Category
Data Exfiltration
Confidence
60% confidence
Finding

Code enumerates, copies, or searches environment variables for secrets. Bulk environment access can collect credentials unrelated to the skill's stated purpose.

Content

Scanner excerpt · tests/test_rd_capitalization.py (reported line 31)May include surrounding context.

python
def run(*args, env=None, cwd=None):
    e = dict(os.environ, PYTHONDONTWRITEBYTECODE="1")
    if env:
        e.update(env)
    return subprocess.run([sys.executable, "-B", str(args[0]), *map(str, args[1:])],

Env Variable Harvesting

High
Category
Data Exfiltration
Confidence
60% confidence
Finding

Code enumerates, copies, or searches environment variables for secrets. Bulk environment access can collect credentials unrelated to the skill's stated purpose.

Content

Scanner excerpt · tests/test_rd_capitalization.py (reported line 134)May include surrounding context.

python
os.chmod(td, 0o700)

    def test_no_home_is_silent(self):
        e = dict(os.environ, PYTHONDONTWRITEBYTECODE="1")
        e.pop("HOME", None)
        r = subprocess.run([sys.executable, "-B", str(CHECK), "--demo"],
                           capture_output=True, text=True, env=e, cwd=str(PKG))

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

Authenticated tool discovery is legitimate, but the surrounding instructions derive endpoint routing from a live external catalog without specifying credential-recipient validation. If that catalog or endpoint configuration is compromised, following this workflow can send CUE_API_KEY to an unauthorized server.

Content

Scanner excerpt · README.en.md (reported line 106)May include surrounding context.

md
curl -s https://cuecue.cn/api/mcp-catalog | head -c 400

# 2. once connected: list the tools, trust the live return
curl -s -X POST "$MCP_ENDPOINT" \
     -H "Authorization: Bearer $CUE_API_KEY" \
     -H 'Content-Type: application/json' \
     -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | head -c 800

The workflow invokes a third-party document parser without warning that source-document content may be sent outside the user's environment.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The workflow initiates an external parsing channel for attachments without explicitly explaining the document-disclosure boundary or requiring authorization for that transfer. Annual reports are normally public, reducing typical impact, but user-supplied files can contain confidential material or annotations.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

The workflow combines live catalog-controlled routing with a request carrying CUE_API_KEY, without requiring independent validation of the credential recipient. A compromised catalog or altered endpoint configuration could therefore redirect the bearer credential to an unauthorized destination.

Content

Scanner excerpt · README.md (reported line 106)May include surrounding context.

md
curl -s https://cuecue.cn/api/mcp-catalog | head -c 400

# 二、已连上后取工具清单:以现场返回为准
curl -s -X POST "$MCP_ENDPOINT" \
     -H "Authorization: Bearer $CUE_API_KEY" \
     -H 'Content-Type: application/json' \
     -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | head -c 800

External Transmission

Medium
Category
Data Exfiltration
Confidence
91% confidence
Finding

The skill tells the agent to obtain domains from a live catalog and then authenticate to the resulting endpoint, without an independent destination check. If routing is maliciously changed, this crosses a credential trust boundary and can expose the API key; the tools/list payload itself is not sensitive document content.

Content

Scanner excerpt · SKILL.en.md (reported line 64)May include surrounding context.

md
curl -s https://cuecue.cn/api/mcp-catalog | head -c 400

  # 2. once connected: list the tools, trust the live return
  curl -s -X POST "$MCP_ENDPOINT" \
       -H "Authorization: Bearer $CUE_API_KEY" \
       -H 'Content-Type: application/json' \
       -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | head -c 800

The attachment-parsing workflow lacks a warning that user-provided documents will be sent to an external service.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

Attachments and scans are directed through a server-backed parsing channel without an explicit warning about document transmission or a confidentiality check. Public annual reports limit ordinary exposure, but the same trigger accepts user-provided documents that may not be public.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env, file_read, file_write, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
91% confidence
Finding

Live catalog routing determines the service destination, while the subsequent request sends a bearer API key without a stated independent origin-validation requirement. A compromised catalog or manipulated endpoint can therefore turn legitimate discovery into credential disclosure.

Content

Scanner excerpt · SKILL.md (reported line 82)May include surrounding context.

md
curl -s https://cuecue.cn/api/mcp-catalog | head -c 400

  # 二、已连上后取工具清单:以现场返回为准
  curl -s -X POST "$MCP_ENDPOINT" \
       -H "Authorization: Bearer $CUE_API_KEY" \
       -H 'Content-Type: application/json' \
       -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | head -c 800

The attachment-parsing workflow lacks a warning that document content may be sent to a third-party service.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The attachment trigger starts a third-party parsing channel without clearly stating that document content may leave the user's environment or checking authorization. The public-disclosure focus reduces normal sensitivity, but does not protect nonpublic user uploads or annotated copies.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · tests/test_rd_capitalization.py (reported line 34)May include surrounding context.

python
e = dict(os.environ, PYTHONDONTWRITEBYTECODE="1")
    if env:
        e.update(env)
    return subprocess.run([sys.executable, "-B", str(args[0]), *map(str, args[1:])],
                          capture_output=True, text=True, env=e, cwd=str(cwd or PKG))

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · tests/test_rd_capitalization.py (reported line 136)May include surrounding context.

python
def test_no_home_is_silent(self):
        e = dict(os.environ, PYTHONDONTWRITEBYTECODE="1")
        e.pop("HOME", None)
        r = subprocess.run([sys.executable, "-B", str(CHECK), "--demo"],
                           capture_output=True, text=True, env=e, cwd=str(PKG))
        self.assertEqual(r.returncode, 0)
        self.assertEqual(r.stderr, "", f"无 HOME 不许吐杂音:{r.stderr[:200]}")

The documented first-run usage beacon creates persistent files unrelated to R&D disclosure analysis.

Low
Category
Excessive Agency
Confidence
90% confidence
Finding

The README states that every first run, including any subcommand, creates a dated marker under ~/.cue-usage/ to measure actual package use. This installation-usage tracking is unrelated to extracting, sourcing, calculating, or checking the R&D capitalization register; the marker is local, and no transmission is shown.

Content

No source excerpt is available for this finding.

The documented startup usage tracking creates persistent files unrelated to R&D disclosure analysis.

Low
Category
Excessive Agency
Confidence
95% confidence
Finding

The README states that running any subcommand creates a dated marker under ~/.cue-usage/ to establish that the skill was installed and run. This installation-usage tracking is unrelated to extracting, citing, or validating R&D capitalization disclosures and is not part of the manifest's stated purpose.

Content

No source excerpt is available for this finding.

Unrelated usage tracking writes persistent files in the user's home directory.

Low
Category
Excessive Agency
Confidence
99% confidence
Finding

The checker silently creates dated usage markers under ~/.cue-usage before validating its input, including during demo and self-test runs. This unrelated persistent write records skill usage without consent and exceeds the expected behavior of a disclosure checker, although no network transmission or sensitive-content collection is shown.

Content

No source excerpt is available for this finding.

The no-environment-variable guarantee contradicts home-directory resolution.

Low
Category
Excessive Agency
Confidence
98% confidence
Finding

The claimed absence of environment-variable reads is inaccurate: os.path.expanduser("~") normally consults HOME on Unix-like systems, and the resolved directory receives usage markers. This misrepresents the checker's privacy boundary and allows the environment to influence the write location, but does not demonstrate credential harvesting or exfiltration.

Content

No source excerpt is available for this finding.

Tests enforce an undeclared usage-tracking capability unrelated to R&D disclosure analysis.

Low
Category
Excessive Agency
Confidence
96% confidence
Finding

These tests explicitly require persistent usage-marker creation, making unrelated tracking part of the accepted checker contract. The cited tests redirect HOME to temporary directories, limiting their direct impact, but they reinforce the production privacy issue rather than requiring user consent.

Content

No source excerpt is available for this finding.