Back to skill

Security audit

Clinical Trial Registry Search / 临床试验注册检索专家

Security checks for vulnerabilities and agentic risk

Overview

ClawScan rates this skill suspicious with high confidence because scanners report substantial undeclared behavior, data-sharing and file-handling risks, and a few insecure practices; no evidence here establishes malicious intent, and coverage is incomplete.

Treat this as a review-required skill. Before installing, consider whether its document-processing and extra data-sharing features fit your needs, whether you accept the reported cleanup and file-path risks, and whether you can avoid sending sensitive text or using the affected network features. The scan covered 79 of 85 components fully, with six partially inspected, so it cannot 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/export_xlsx.py
Finding

External registry text is interpreted as Excel formulas

Content
View full analysis

Impact

An attacker-controlled field can calculate misleading report values or introduce deceptive formula-generated links. Additional effects depend on the spreadsheet application's supported functions and security settings; credential theft, successful network exfiltration, and operating-system code execution were not demonstrated.

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/ct_registry.py
Finding

Default cleanup deletes pre-existing files unrelated to the current run

Content
View full analysis

Impact

Existing notes, JSON datasets, or configuration files directly inside the output directory can be permanently deleted. If deletion fails, cleanup attempts to move them into _unsaved instead.

Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • 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 (143)

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · CHANGELOG.md (reported line 741)May include surrounding context.

md
## v0.9.3 (2026-09-06) · 新增本地 Web 检索工作台 workbench/ (v0.1.1, clinical-indigo)

- **新增 `workbench/` 目录**:本地可运行的单页检索控制台。前端 `index.html`(单文件、全内联零外链、SVG 图标、zh/en 97 键 1:1 对齐)+ 后端 `server.py`(stdlib `http.server` 编排 `scripts/ct_registry.py` 子进程,默认端口 8790,已加 CORS+PNA;端点 `/api/config /api/health /api/search /api/runs /api/task /api/download /api/run-delete /api/bug-report`)。
- **品牌统一 clinical-indigo**:与技能 `assets/icon.svg` 一致(`#1C6DD0`);顶栏 logo 字节级内联真实 `icon.svg`(含 `logoShadow` 投影滤镜 + `translate(83,40)` 十字医学标记),去除此前自绘变体。
- **发布前修复(A/B/C/D/G 五项)**:1 配置漂移 `bugReport.enabled` 在 `workbench.config.json` 与 `#wb-config` 兜底双处统一为 `true`;2 `server.py` 去掉硬编码绝对路径,改为 `_resolve_skill_dir()`(`CT_REGISTRY_SKILL` 环境变量 > 脚本父目录 > 向上回溯兜底)+ `PY=sys.executable`;3 logo 替换为真实 `icon.svg`、品牌切 clinical-indigo;4 兜底 `modules` 补齐为 4 区(search-form/result/tracking/run-history);5 `run_workbench.cmd` 改可移植 `python server.py`。
- **质量闸门**:ct-base §9 质量闸门 10/10 PASS;§16 三道闸门(secret_scan / shared_sync_check / spec_lint)对 workbench

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
80% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · CHANGELOG.md (reported line 741)May include surrounding context.

md
## v0.9.3 (2026-09-06) · 新增本地 Web 检索工作台 workbench/ (v0.1.1, clinical-indigo)

- **新增 `workbench/` 目录**:本地可运行的单页检索控制台。前端 `index.html`(单文件、全内联零外链、SVG 图标、zh/en 97 键 1:1 对齐)+ 后端 `server.py`(stdlib `http.server` 编排 `scripts/ct_registry.py` 子进程,默认端口 8790,已加 CORS+PNA;端点 `/api/config /api/health /api/search /api/runs /api/task /api/download /api/run-delete /api/bug-report`)。
- **品牌统一 clinical-indigo**:与技能 `assets/icon.svg` 一致(`#1C6DD0`);顶栏 logo 字节级内联真实 `icon.svg`(含 `logoShadow` 投影滤镜 + `translate(83,40)` 十字医学标记),去除此前自绘变体。
- **发布前修复(A/B/C/D/G 五项)**:① 配置漂移 `bugReport.enabled` 在 `workbench.config.json` 与 `#wb-config` 兜底双处统一为 `true`;② `server.py` 去掉硬编码绝对路径,改为 `_resolve_skill_dir()`(`CT_REGISTRY_SKILL` 环境变量 > 脚本父目录 > 向上回溯兜底)+ `PY=sys.executable`;③ logo 替换为真实 `icon.svg`、品牌切 clinical-indigo;④ 兜底 `modules` 补齐为 4 区(search-form/result/tracking/run-history);⑤ `run_workbench.cmd` 改可移植 `python server.py`。
- **质量闸门**:ct-base §9 质量闸门 10/10 PASS;§16 三道闸门(secret_scan / shared_sync_check / spec_lint)对 workbench

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · README.md (reported line 195)May include surrounding context.

md
WHO ICTRP and China CDE **share one long-lived credential** on the unified endpoint — it is a public shared credential that ships with the skill, so it works out-of-the-box. No confidential data ever reaches any of these endpoints.

### Access credential

The unified endpoint needs a shared access credential, which **ships with the skill** — retrieval runs out of the box with no setup, no key to apply for, and nothing for you to paste. If the author ever re-issues it, you will be told what to do; never paste credentials into chat.

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · README.md (reported line 197)May include surrounding context.

md
### Access credential

The unified endpoint needs a shared access credential, which **ships with the skill** — retrieval runs out of the box with no setup, no key to apply for, and nothing for you to paste. If the author ever re-issues it, you will be told what to do; never paste credentials into chat.

## Why You Can Trust the Results (anti-hallucination)

Credential Access

High
Category
Privilege Escalation
Confidence
70% confidence
Finding

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

Content

Scanner excerpt · README.md (reported line 222)May include surrounding context.

md
|---|---|---|
| CT.gov connection error | Network or proxy unreachable | Check clinicaltrials.gov is reachable, or set your proxy |
| CDE / ChiCTR "access blocked" | The site's firewall rejects automated browsers | Use the unified-endpoint route (the default); if it persists, the assistant falls back to an assisted paste that sends nothing |
| HTTP 401 on WHO / CDE | Access credential not accepted | Retry; if it persists, contact the author for a re-issued credential |
| HTTP 403 (rare) | Corrupted credential | Contact the author; this is not an expiry |
| CDE HTTP 500 field-type error | A structured field was sent in the wrong shape | Re-run with the plain form; the assistant rebuilds the request |
| ISRCTN returns 0 rows / anomalous | The official XML API returned a soft error | The orchestration auto-switches to the WHO mirror route to filter ISRCTN records |

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
100% confidence
Finding

The supplied chunk's primary purpose is Word-comment inspection and modification, not clinical-trial registry retrieval or aggregation. Its document reads and DOCX writes also exceed the declared filesystem scope, which limits reads to the skill's own files and specifies report/Excel writes in the current working directory, with a separate PDF-download exception. These are substantive undeclared capabilities, not supporting details of registry search. No network transmission is present, and the code does not enforce the comment's de-identification warning.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Automatic online translation introduces additional recipients for search terms, while the network declaration lists registry retrieval services rather than translation providers. The body does disclose MyMemory and a disable flag, reducing concealment concerns, but the reported Google destination and recipient expansion still require explicit disclosure; cache writes alone are not a security exploit.

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 the absence of registry-search logic in this helper module is not itself a mismatch. However, the declared filesystem permission explicitly makes the skill's own files read-only, while the code supports modifying configuration and session files there. These writes are outside the declared report/Excel and optional PDF output exceptions. The mutation functions must be invoked; they do not execute merely upon import.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

UI validation can reasonably support a clinical-trial search workbench, so the absence of registry-search logic in this helper is not itself a mismatch. However, its explicit repair modes modify application assets and skill metadata, and its multi-skill scan reads sibling skill resources. Those capabilities exceed the stated filesystem permissions. These are directly implemented behaviors, not merely instructions or examples in comments. No network transmission or confidential-data collection is evident in this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied module implements clinical dataset submission-metadata tooling rather than registry search or trial-landscape reporting. Its main entry points, build_define_xml and validate_define_xml, generate and inspect CDISC XML; no registry retrieval, trial aggregation, or JSON/Markdown report production appears in this chunk. Although it may be a shared library, its substantive capabilities are not described by the declared purpose. The shown implementation contains no direct network requests or apparent exfiltration.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Sending full messages and action parameters to configurable LLM endpoints materially exceeds the promise to transmit only public search terms and limited metadata. Although the body mentions possible LLM calls generically, that does not authorize full-message transmission, and endpoint configuration can also expose API credentials to an unintended recipient.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The reported addition of account_id and billing_token sends identity and potentially sensitive authentication material beyond the enumerated metadata contract. Billing integration may be legitimate, but silently attaching these fields undermines informed consent about outbound data; unrestricted caller-selected output paths are not independently a vulnerability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The reported deletion or relocation of pre-existing Markdown and JSON files without checking ownership can destroy unrelated user work in a shared output directory. The vulnerability is the unscoped destructive cleanup, not merely support for configurable output paths.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The reported transmission of unrestricted problem descriptions and diagnostics exceeds the public-query-only egress boundary. In a clinical workflow, those fields can contain patient information, internal paths, credentials, or unpublished study details; the arbitrary local output directory alone is not evidence of exploitation.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Quota enforcement is a reasonable supporting implementation detail for registry retrieval and is not itself a purpose mismatch. However, the explicit filesystem permissions do not disclose its state and lockfile writes. By default these occur under ~/.workbuddy/skills/ct-registry/config/, rather than as reports in the working directory or requested document downloads. The environment override also permits other writable locations. The mismatch is therefore in filesystem behavior, not evidence of unrelated retrieval or data exfiltration.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The mismatch is not simply that this helper lacks registry-search functionality; supporting utilities need not implement the entire skill. The concrete discrepancy is that it accepts and reads arbitrary supplied Office-file paths despite the explicit own-files-only read permission. Its broad document-content extraction is also absent from the stated public-query-only workflow. External URLs are extracted as text, not fetched, and this chunk provides no evidence of data transmission or exfiltration.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The reported upload of locally loaded project_list records crosses the public-query-only boundary and could disclose unpublished trial or sponsor information. Confidence is reduced because the implementation is not supplied and SKILL.md identifies the standalone CDE workflow as retired and excluded from the package, so deployed reachability remains unverified.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

Generating a machine hash is explicitly disclosed and can support the declared query_origin metadata, so the helper's lack of clinical-trial searching is not itself a mismatch. However, its Linux implementation reads system files outside the expressly permitted filesystem scope. The code also obtains persistent platform identifiers on Windows and macOS, but the supplied chunk does not establish that any identifier is transmitted.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The script is related to clinical-trial planning, but its trial-versus-disease-burden analysis extends beyond the described registry aggregation. More concretely, its input and output path handling exceeds the explicitly declared filesystem boundaries: --trials and --dalys can read files outside the skill directory, while --out and --chart can write outside the current working directory. These operations are user-directed, and there is no evidence of network exfiltration or covert file discovery. The absence of registry-search functionality in this individual supporting script is not itself treated as a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The search, aggregation, and deduplication behavior supports the declared purpose. However, the filesystem behavior exceeds the stated permission boundary: --out accepts an unrestricted path, --workdir can select another directory, and os.makedirs creates that directory if needed. The declared exception for writing outside the current directory applies only to EU-CTR PDFs, not these JSON files. The supplied chunk does not establish what network data the delegated search_ictrp.py transmits.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

Drug-name normalization is a reasonable supporting implementation detail for registry searches; this helper need not implement the full declared workflow. The mismatch is specifically the filesystem boundary: when the local mapping is absent and the shared mapping exists, the code reads an external ct-base resource. This is a narrow permission discrepancy, not evidence of exfiltration or an unrelated primary purpose.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
92% confidence
Finding

The registry retrieval and normalization behavior matches the declared primary purpose. However, the filesystem permissions are materially understated: this adapter accepts arbitrary output and cache paths and writes there with the process's permissions. The declared exception for user-selected output directories applies only to download_docs.py and retrieved PDFs, not this JSON adapter. Caching itself is a supporting implementation detail; the mismatch is its unrestricted filesystem destination. The code also disables TLS certificate verification, which is a security weakness rather than a separate purpose mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The public PubChem requests and enrichment functionality support the declared purpose. However, the declared filesystem permission limits this adapter's report writes to the current working directory. The code passes --out directly to open(..., 'w') without restricting its location, including in error-result paths. This enables writes and overwrites elsewhere within the executing user's filesystem permissions. No confidential-data collection or unrelated network destination is evident.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

Snapshot comparison is a reasonable supporting capability for the declared trial-analysis purpose, and the lack of registry retrieval in this helper is not itself a mismatch. However, its filesystem behavior exceeds the explicitly declared scope: open() and shutil.copy() accept unrestricted input and output paths, allowing reads and writes outside the stated boundaries. These operations require caller-supplied paths; the code shows no covert collection or network transmission.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The supplied script implements a generic workbench UI linter, not registry searching or trial-data processing. UI validation could be legitimate supporting tooling, so the absence of registry functionality in this helper alone is not decisive. However, its explicit path arguments permit reading HTML and rule files, and recursively discovering entry files, outside the skill directory, contrary to the declared own-files-only read restriction. The containment check for linked assets confines them to the selected HTML directory, not to the skill directory. No network transmission, file writing, or code execution is shown.

Content

No source excerpt is available for this finding.