Back to skill

Security audit

Markdown 转 Word · MD 转 DOCX(表格还原)

Security checks for vulnerabilities and agentic risk

Overview

The skill is rated suspicious because scanners report remote images that may trigger network requests, unpinned installation guidance, and several reliability and disclosure concerns; the scan was partial.

Consider whether you are comfortable with a skill whose generated HTML may contact external image hosts and whose installation guidance uses unpinned upstream content. Review and pin the installation sources, and avoid using it with sensitive documents until you have checked its network behavior and the reported output-validation limitations. Coverage was marked partial because some local references 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
  • 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
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
Findings (22)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
88% confidence
Finding

The HTML rendering behavior closely matches the declared purpose, and CSS inlining is a supporting implementation detail. However, PDF generation is a distinct, user-selectable output capability absent from the description. The supplied chunk need not implement Word or PPT itself to be consistent with a broader multi-file skill. No unrelated data collection or exfiltration is evident.

Content

No source excerpt is available for this finding.

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · scripts/render_html.py (reported line 429)May include surrounding context.

python
if one and ":" not in one and decls:
                    rules.append((one, decls))
        i = j
    return rules


_COMPOUND_RE = re.compile(r"^(?:([a-zA-Z][\w-]*))?((?:\.[\w-]+)*)?(?:#([\w-]+))?$")

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
70% confidence
Finding

npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

The installation command executes an unpinned npm CLI and installs skill content from a mutable repository reference. Although this is not itself an MCP server declaration, it creates a supply-chain exposure: compromised upstream releases or repository changes could introduce code or agent instructions with the user's privileges.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

The optional parsing-channel setup uses an unpinned npm installer and mutable external skill content. Upstream compromise could therefore affect the host during installation or subsequent agent use; the credential-storage guidance does not address this separate supply-chain risk.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

The command invokes an unpinned npm CLI and obtains skill content without an immutable revision. This is an installation supply-chain weakness rather than an unpinned MCP server configuration, but compromised upstream content could execute locally or influence the agent.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

The reader-skill installation depends on mutable npm and repository content. Parsing is optional, which limits exposure, but users following this setup could receive compromised installer code or skill instructions before configuring the channel.

Content

No source excerpt is available for this finding.

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

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
92% confidence
Finding

Agent-facing setup guidance recommends executing an unpinned npm CLI and installing a reader skill from mutable upstream content. Although the detector mislabels this as an MCP server reference, the supply-chain exposure is real and is more actionable because an agent may follow the installation guidance.

Content

No source excerpt is available for this finding.

The PDF-export workflow lacks a warning that remote image links can trigger external network requests.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The agent-facing PDF workflow does not warn that source-controlled remote image URLs can cause the browser to make network requests. This can enable tracking or requests to internal services reachable by the rendering host; disclosure in the READMEs mitigates transparency concerns but does not ensure the agent checks this boundary before export.

Content

No source excerpt is available for this finding.

The HTML output is documented as self-contained but can depend on remote images.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Despite the self-contained-output claim, HTTP and HTTPS image URLs are embedded directly in the generated HTML. This breaks the implied offline/privacy boundary: opening or exporting an attacker-supplied document can contact external servers, overlapping with the network-request finding below.

Content

No source excerpt is available for this finding.

Remote images can trigger undisclosed network requests when viewing or exporting the generated HTML.

Medium
Category
Output Handling
Confidence
98% confidence
Finding

Untrusted Markdown image URLs become automatically loaded browser resources, including during headless PDF export. An attacker can use unique URLs to track document access and disclose the viewer's network address, or induce requests to reachable internal endpoints; the code does not demonstrate arbitrary local-file exfiltration.

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 · scripts/render_html.py (reported line 344)May include surrounding context.

python
return False, "未找到 headless 浏览器(Edge/Chrome/Chromium),跳过 PDF"
    url = "file:///" + os.path.abspath(html_path).replace("\\", "/")
    try:
        subprocess.run([browser, "--headless", "--disable-gpu", "--no-pdf-header-footer",
                        "--print-to-pdf=%s" % os.path.abspath(pdf_path), url],
                       stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, timeout=120)
    except Exception as e:

Documented prose-exclusion and slide-limit options are parsed but not passed to the renderer.

Medium
Category
Excessive Agency
Confidence
100% confidence
Finding

Both renderer call paths omit the parsed prose and slide-limit options, leaving the defaults of storing prose in speaker notes and allowing up to 80 slides. A user selecting prose exclusion can therefore distribute sensitive text hidden in the PPTX notes, while a requested slide cap is also ignored.

Content

No source excerpt is available for this finding.

The validator accepts multi-source input but ignores the source chapters and analysis when assessing fidelity.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The JSON validation path reads only the top-level markdown field and ignores source chapters and integrated analysis. Multi-source input without that field produces an empty reference and defaults to 100% coverage, allowing incomplete output to receive a misleading fidelity pass.

Content

No source excerpt is available for this finding.

The claimed near-zero-text-loss validation measures distinct CJK character presence, not preservation of source text.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The pass/fail fidelity check compares sets of distinct CJK characters, discarding order, repetition, and non-CJK content. Substantially deleted or altered documents can consequently pass, undermining this skill's advertised fidelity gate, particularly for financial or legal text containing important numbers.

Content

No source excerpt is available for this finding.

The documented acceptance criterion is CJK coverage ≥99%, but unrelated visual-design checks can also fail validation.

Medium
Category
Excessive Agency
Confidence
95% confidence
Finding

The module documentation at L008–L010 states that the verdict is based on CJK character coverage and describes structural checks as informational. However, L180–L189 add rounded corners, shadows, color count, missing CJK fonts, and contrast to the failure conditions, and L190–L192 use these conditions to determine the verdict and exit status. Consequently, an HTML document with complete source-text coverage can fail despite satisfying the stated acceptance criterion.

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/run_all.py (reported line 23)May include surrounding context.

python
print("\n" + "#" * 72)
        print("# SUITE:", suite)
        print("#" * 72)
        rc = subprocess.run([PY, str(HERE / suite)]).returncode
        if rc != 0:
            failed.append(suite)
    print("\n" + "=" * 72)

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/run_escape_test.py (reported line 54)May include surrounding context.

python
md = OUT / (name + ".md")
    htm = OUT / (name + ".html")
    md.write_text(md_text, encoding="utf-8")
    r = subprocess.run([PY, "-B", RENDER, "--md", str(md), "--out", str(htm)],
                       capture_output=True, text=True)
    if r.returncode != 0 or not htm.is_file():
        raise SystemExit("3 取值域故障:render_html 没出件(%s|rc=%d|%s)"

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/run_matrix_multi.py (reported line 42)May include surrounding context.

python
def run(cmd):
    return subprocess.run([PY] + cmd, capture_output=True, text=True,
                          encoding="utf-8", errors="replace")

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/run_matrix_test.py (reported line 155)May include surrounding context.

python
docx_out = OUT / f"t_{case_id}.docx"
        cmd = [PY, BUILD, "--json", str(tmp_json), "--out", str(docx_out),
               "--profile", profile] + extra
        r = subprocess.run(cmd, capture_output=True, text=True,
                           encoding="utf-8", errors="replace")
        rec["build_rc"] = r.returncode
        if r.returncode != 0:

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/run_matrix_test.py (reported line 164)May include surrounding context.

python
results.append(rec)
            print(f"[BUILD FAIL] {case_id}: {rec['err']}")
            continue
        v = subprocess.run([PY, VALID, "--json", str(tmp_json), "--docx", str(docx_out)],
                           capture_output=True, text=True, encoding="utf-8",
                           errors="replace")
        val_ok = v.returncode == 0