Back to skill

Security audit

Paper Polisher Pro — AI Detector & Academic Polishing

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because scanners report cross-origin access to local MCP tools and several mismatches between its advice-only description and reported editing or verdict behavior; the review is incomplete.

Consider the suspicious verdict before installing. Review the local MCP server exposure and advice-only scope mismatches, and note that coverage is partial: 177 of 180 components were scanned, with three partially inspected and a security-relevant parser limit. The results do not establish malicious intent or a compromised dependency, but they also do not establish that the skill is safe.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/pp_mcp_server.py
Finding

Cross-origin requests can invoke local MCP tools

Content
View full analysis

Impact

Unauthorized invocation of analysis and diagnostic tools consumes local CPU and serializes legitimate clients behind attacker-selected work. Workflow calls also create report directories that tool_workflow() does not remove.

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/pp_mcp_server.py
Finding

One incomplete request can indefinitely block all HTTP clients

Content
View full analysis

Impact

All subsequent MCP requests and health checks remain unavailable until the offending connection closes or the server is restarted. The body-size ceiling does not prevent an incomplete small request from blocking.

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (171)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Most of this script supports the declared terminology-checking purpose and local-only operation. However, main() invokes auto_fix() when --auto-fix is supplied and issues exist, then writes the modified text to a sibling file with '_fixed' added to its stem. This is an executable rewriting capability, not merely a suggestion. Moreover, auto_fix() searches for every occurrence of each flagged original string rather than limiting changes to the reported positions, so replacements can extend beyond the specific occurrences that were flagged. The broader skill features absent from this individual code chunk are not treated as mismatches.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The gate contains no input-length check before scoring or assigning a verdict. Even when the style layer reports that text is too short, layer4_style substitutes a neutral score of 50, and _gate_impl still produces a categorical verdict. Failed or unparseable scoring layers likewise receive fallback scores rather than causing abstention. This contradicts the explicit promise that texts under 100 characters receive no verdict. Other advertised features not present in this individual code chunk are not, by themselves, evidence of a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

extract_features rejects only texts shorter than 50 characters and requires at least two qualifying sentence segments. style_distance applies no additional 100-character guard before issuing a verdict, so qualifying inputs of 50–99 characters receive classifications. Otherwise, this chunk is consistent with a supporting local style-analysis component; it contains no network access or automatic rewriting. Broader features absent from this individual module are not treated as mismatches.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The comparison mode explicitly rates repetition-reduction effectiveness as limited, moderate, or good based on average 2/3/4-gram overlap, and may recommend further rewriting. This is a distinct evaluation capability not disclosed in the description, although internal repetition analysis could support writing-quality checks. Neither mode checks text length: even empty or very short inputs receive categorical assessments, directly contradicting the declared no-verdict rule. The script does not execute rewrites, perform corpus-based plagiarism detection, or access the network; its local file reads and optional report writing are not themselves mismatches. Other advertised features being absent from this individual code chunk is not treated as evidence of mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The evaluation setting makes adversarial sample generation relevant to testing the detector, but the script implements and executes the very rewriting capability the description categorically excludes. Its paraphrase function modifies text automatically, and main persists those modifications as generated corpus records. This is a narrow capability mismatch, not evidence of an unrelated or malicious purpose. The file access is local and consistent with the zero-upload claim; no network activity is present in this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

Before-and-after comparison is a supporting capability consistent with the declared self-checking purpose, not automatic rewriting. However, the description promises that texts under 100 characters receive no verdict by design. main() enforces that limit, while rewrite_check() checks only that inputs are nonempty strings and then unconditionally processes detector scores into a comparison report. Thus the supplied code does not enforce the declared boundary for programmatic use. Imported detector behavior is not supplied, so any additional short-text protection inside that dependency cannot be verified.

Content

No source excerpt is available for this finding.

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 138)May include surrounding context.

md
ly (what not to do); operational how-to questions live in [`references/faq.md`](references/faq.md) (searchable via `pp.py faq <keyword>`) — the two cross-refere

Incomplete referenced artifact analysis

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 256)May include surrounding context.

md
## FAQ (top 3 — full grouped FAQ in [`references/faq.md`](references/faq.md), searchable via `pp.py faq <keyword>`)

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 · eval/release_smoke.py (reported line 54)May include surrounding context.

python
def run_py(script, args, env_extra=None, timeout=180):
    env = dict(os.environ)
    if env_extra:
        env.update(env_extra)
    p = subprocess.run([sys.executable, str(SCRIPTS / script)] + args,

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_v60000.py (reported line 22)May include surrounding context.

python
def _run(*args, timeout=180, env_extra=None):
    env = dict(os.environ)
    if env_extra:
        env.update(env_extra)
    return subprocess.run([PY, *map(str, args)], capture_output=True,

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_v60100.py (reported line 21)May include surrounding context.

python
def _run(*args, timeout=120, env_extra=None):
    env = dict(os.environ)
    if env_extra:
        env.update(env_extra)
    return subprocess.run([PY, *map(str, args)], capture_output=True,

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_v60300.py (reported line 18)May include surrounding context.

python
def _run(*args, timeout=60, env_extra=None):
    env = dict(os.environ)
    if env_extra:
        env.update(env_extra)
    return subprocess.run([PY, *map(str, args)], capture_output=True,

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 · eval/release_smoke.py (reported line 646)May include surrounding context.

python
_p = subprocess.run([sys.executable, str(SCRIPTS / "ai_detector.py"), _hu_file_hold,
                         "--profile", "journal", "--format", "json"],
                        capture_output=True, text=True, encoding="utf-8", timeout=180,
                        env={**os.environ, "PP_ORT_THREADS": "8"})
    try:
        _jd = json.loads(_p.stdout)
        _ok_j = _jd.get("journal_precheck") is not None and _jd.get("risk_bands") is not None

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 · eval/release_smoke.py (reported line 674)May include surrounding context.

python
_pp_d = subprocess.run([sys.executable, str(SCRIPTS / "pp.py"), "detect", _route_f,
                            "--format", "json"], capture_output=True, text=True,
                           encoding="utf-8", timeout=180,
                           env={**os.environ, "PP_ORT_THREADS": "8"})
    Path(_route_f).unlink(missing_ok=True)
    try:
        _route_ok = _pp_d.returncode == 0 and isinstance(json.loads(_pp_d.stdout).get("overall_ai_score"), (int, float))

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 · eval/release_smoke.py (reported line 689)May include surrounding context.

python
try:
    _p = subprocess.run([sys.executable, str(SCRIPTS / "pp_setup.py"), "--check"],
                        capture_output=True, text=True, encoding="utf-8", timeout=180,
                        env={**os.environ, "PP_ORT_THREADS": "8"})
    chk("pp_setup --check 可跑", _p.returncode in (0, 1), f"rc={_p.returncode}")
    _fake = Path(tempfile.mkdtemp()) / "fake.onnx"
    _fake.write_bytes(b"not a model")

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 · eval/release_smoke.py (reported line 843)May include surrounding context.

python
# 版本测试文件自动 +75s——不再每次撞墙手调。基线 420s + 每文件 75s。
    _ntestfiles = len(list((ROOT / "tests").glob("test_*.py")))
    _ut_timeout = 420 + 75 * _ntestfiles
    _ut_env = dict(os.environ, PP_IN_SMOKE="1")  # v7.0.0: 嵌套标记(防 smoke→单测→smoke 套娃)
    _ut = subprocess.run([sys.executable, "-m", "unittest", "discover", "-s", "tests",
                          "-t", "."], capture_output=True, text=True, encoding="utf-8",
                         cwd=str(ROOT), timeout=_ut_timeout, env=_ut_env)

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 · scripts/pp_mcp_server.py (reported line 61)May include surrounding context.

python
def _run_script(args, timeout=300, stdin_text=None):
    """跑包内脚本,返回 (rc, stdout, stderr)。UTF-8 强制。"""
    env = dict(os.environ, PYTHONIOENCODING="utf-8")
    p = subprocess.run([sys.executable] + args, capture_output=True,
                       timeout=timeout, env=env, input=stdin_text, text=True,
                       encoding="utf-8", errors="replace")

The local-listener exemption checks source substrings rather than the promised loopback binding.

High
Category
Excessive Agency
Confidence
100% confidence
Finding

The verifier removes http.server findings whenever two substrings appear anywhere in the source, including comments or unused strings, without checking the actual listener address. A non-loopback server can therefore pass this exemption, potentially exposing manuscript-related services while users rely on the skill's security self-verification; this is a verification bypass, not evidence that exposure currently occurs.

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_v60400.py (reported line 120)May include surrounding context.

python
self.assertIn("路线图小结", p.stdout)

    def test_tour_en_runs(self):
        env = dict(os.environ, PP_LANG="en")
        p = subprocess.run([sys.executable, str(SCRIPTS / "pp.py"), "tour"],
                           capture_output=True, text=True, timeout=300, env=env)
        self.assertEqual(p.returncode, 0, p.stderr[-300:])

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_v60700.py (reported line 71)May include surrounding context.

python
self.assertIn("投稿前打包自查", r.stdout)

    def test_en_surface(self):
        env = dict(os.environ, PP_LANG="en")
        r = _run([str(SCRIPTS / "pp.py"), "scenarios"], timeout=120)
        r_en = subprocess.run([sys.executable, str(SCRIPTS / "pp.py"), "scenarios"],
                              capture_output=True, text=True, timeout=120, env=env)

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.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

The documented 50/50 mixed-text generator can produce entirely human text labeled as an AI-containing attack.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

When the smaller paragraph pool has one paragraph, mix_doc selects only the human paragraph, and the caller can persist it with label=1 and attack=mixed if it meets the length check. This corrupts evaluation ground truth and can distort the robustness measurements used to justify detector reliability.

Content

No source excerpt is available for this finding.

Unconditional shuffling destroys the benchmark's documented paragraph-order scenarios.

Medium
Category
Excessive Agency
Confidence
100% confidence
Finding

Unconditional shuffling removes the ordering guarantees of interleave, ai_lead, and human_lead while retaining those scenario labels. Paragraph truth labels remain aligned, but scenario-specific evaluation no longer measures the documented conditions and can conceal order-dependent weaknesses.

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.