Back to skill

Security audit

长文档全文翻译流水线

Security checks for vulnerabilities and agentic risk

Overview

The skill is rated suspicious with medium confidence: scanners report an unpinned install command, a mismatch between its stated translation purpose and packaging code, and other workflow and privacy concerns; inspection was partial.

Consider holding off on installation until the installer and skill revision can be pinned to reviewed versions. If you still choose to install, review the source and packaging mismatch, and check whether documents may be sent to an external parsing service. Inspection covered 24 of 25 components, with one partially inspected, 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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
Findings (14)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied code has a materially different primary purpose: preparing a skill distribution archive. It does not process foreign-language articles or produce any of the four declared editorial deliverables. Although packaging can be ancillary tooling, this chunk exclusively implements a release workflow rather than supporting the declared text-processing behavior directly. Its safety scan only warns; detected patterns do not block packaging. No network requests, automatic uploads, or execution of scanned code are present.

Content

No source excerpt is available for this finding.

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
95% confidence
Finding

Although this is an npm installer rather than an MCP server, the recommended command executes skills without pinning its version and installs skill content without specifying an immutable revision. A compromised or unexpectedly changed upstream release could therefore execute code or introduce malicious agent instructions under the user's account.

Content

No source excerpt is available for this finding.

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

Medium
Category
MCP Rug Pull
Confidence
95% confidence
Finding

The installation instructions invoke an unpinned npm executable and do not select an immutable revision of the skill repository. This is a supply-chain exposure rather than an MCP-server configuration issue: an upstream compromise could execute code or install malicious skill instructions.

Content

No source excerpt is available for this finding.

The parsing workflow lacks a privacy warning about sending source documents to the omni-reader service.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The workflow recommends an omni-reader parsing service without explaining whether source documents leave the host or requiring approval before external processing. For potentially confidential books or archives, an externally hosted parser could disclose entire documents; authorized access to a document does not itself authorize third-party disclosure.

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, network, 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.

The documented blocking security check only warns and still produces a release archive.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

The release tool promises to abort on failed security checks, but detected risky patterns only generate warnings and the archive is still produced. This fail-open release gate can allow suspicious code into a distributed skill despite the documented safeguard; a match alone does not establish that the code is malicious.

Content

No source excerpt is available for this finding.

Tainted flow: 'OUT_HTML' from os.environ.get (line 23, credential/environment) → open (file write)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/build_reader.py (reported line 294)May include surrounding context.

python
body.count('class="fnref"'),
                        body.count('class="pagemark"'),
                        n_terms, n_doubts)))
    open(OUT_HTML, "w", encoding="utf-8").write(html)
    print("reader html:", OUT_HTML, len(html), "chars")
    print("  脚注链接 %d / 页边码锚点 %d" % (body.count('class="fnref"'),
                                            body.count('class="pagemark"')))

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/check_translation.py (reported line 57)May include surrounding context.

python
def run(script, *args):
    r = subprocess.run([sys.executable, '-B', script] + [str(a) for a in args],
                       capture_output=True, text=True)
    return r.returncode, r.stdout + r.stderr

Page-marked duplicates are retained rather than reduced to page markers, contradicting the documented deduplication strategy.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Lines L009–L011 promise to remove duplicate text while preserving only its page marker, and L096 reiterates this intent. However, after appending the marker at L096, execution falls through to L104 and appends the original duplicate line too. With --apply, this preserves the duplicate translation and adds an extra page marker while reporting the case as repaired.

Content

No source excerpt is available for this finding.

A translation note instructs the reviewer to exempt its contents from reporting.

Medium
Category
Prompt Injection
Confidence
86% confidence
Finding

The note claims that because the word “Sdiell” appears in an annotation, it must be exempted and not reported as untranslated residue. This embeds a directive about the reviewer's reporting behavior inside the material being reviewed, attempting to turn a contextual claim into authority to suppress a finding.

Content

No source excerpt is available for this finding.

Tainted flow: 'out_html' from os.environ.get (line 386, credential/environment) → open (file write)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/merge_build.py (reported line 358)May include surrounding context.

python
<body><div class="layout">{nav}<main>
{body}
</main></div></body></html>"""
    open(out_html, "w", encoding="utf-8").write(html)
    return True

Tainted flow: 'md_path' from os.environ.get (line 382, credential/environment) → open (file write)

Medium
Category
Data Flow
Confidence
65% confidence
Finding

Data from a source is assigned to a variable that is later passed to a sink, creating a variable-mediated taint flow.

Content

Scanner excerpt · scripts/merge_build.py (reported line 383)May include surrounding context.

python
else:
        master, n_files, _ = merge(chunks_dir)
    md_path = out_prefix + ".md"
    open(md_path, "w", encoding="utf-8").write(master)
    print(f"✅ 主稿已生成:{md_path}({len(master)} 字符)")
    html_path = out_prefix + ".html"
    if build_html(master, html_path):

The documented character-count fallback for source chunks is not implemented; chunks without page markers are skipped instead.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Chunks without page markers are excluded from overlap comparisons rather than checked using the documented fallback, and an entirely marker-free input exits with status 0 despite being unvalidated. Mixed inputs can also produce a success message while leaving marker-free chunks unchecked, allowing accidental or deliberately marker-free duplicate content to bypass this quality gate. In this long-document translation workflow, this threatens deliverable integrity through duplicated content, although the shown code does not establish code execution or data disclosure.

Content

No source excerpt is available for this finding.