Back to skill

Security audit

Clinical Trial Safety Signal / 临床试验安全信号专家

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this skill suspicious because the scanners report substantial mismatches with its stated limits on data collection, network destinations, and file access, alongside some security and analysis-quality issues.

Treat the suspicious verdict as a reason to review carefully before installing. Consider whether the reported machine-identifier and billing-data flows, extra network destinations, and storage or file access outside the stated boundaries are acceptable to you. Scanner coverage was partial, so these results do not establish that the skill is safe or malicious.

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/report.py
Finding

Retrieved bulletin content can inject active links and HTML attributes

Content
View full analysis

Impact

A crafted destination can introduce HTML event attributes or an executable URL. Upon interaction, attacker-supplied JavaScript can modify or read the rendered report and potentially transmit its contents, subject to browser and preview restrictions. Local operating-system code execution is not established.

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/report_xlsx.py
Finding

Retrieved bulletin titles are interpreted as spreadsheet formulas

Content
View full analysis

Impact

An attacker-supplied title can become a calculated formula rather than literal evidence. This enables misleading calculated content and, in compatible applications permitting external-data functions, outbound requests that can disclose referenced workbook values. Unconditional command execution is not established.

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 (152)

Tainted flow: 'req' from os.environ.get (line 348, credential/environment) → urllib.request.urlopen (network output)

Critical
Category
Data Flow
Confidence
99% confidence
Finding

Authorization checks validate the non-stream endpoint, but the initial request instead uses the independently environment-configurable CT_SEARCH_ENDPOINT_STREAM without validating it. Someone able to influence that setting could redirect the query payload and Authorization bearer token to an unapproved host; this is a concrete endpoint-validation mismatch rather than generic environment-variable exfiltration.

Content

Scanner excerpt · adapters/coze_dispatch.py (reported line 350)May include surrounding context.

python
"""单次 /stream_run 请求 + SSE 解析;返回 workflow_end.output(dict|list)或 None,异常上抛。"""
    req = urllib.request.Request(CT_SEARCH_ENDPOINT_STREAM, data=body,
                                 headers=_stream_headers(), method="POST")
    with urllib.request.urlopen(req, timeout=timeout) as resp:
        return _parse_stream(resp)

Tainted flow: 'headers' from os.environ.get (line 380, credential/environment) → requests.post (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/nmpa_coze.py (reported line 384)May include surrounding context.

python
safe_payload = _sanitize(payload)
    try:
        resp = requests.post(endpoint, headers=headers, json=safe_payload, timeout=timeout)
    except requests.exceptions.ProxyError:
        # ct-base §5.49:系统代理残留 → 绕代理直连重试
        resp = requests.post(endpoint, headers=headers, json=safe_payload, timeout=timeout,

Tainted flow: 'headers' from os.environ.get (line 380, credential/environment) → requests.post (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/nmpa_coze.py (reported line 387)May include surrounding context.

python
resp = requests.post(endpoint, headers=headers, json=safe_payload, timeout=timeout)
    except requests.exceptions.ProxyError:
        # ct-base §5.49:系统代理残留 → 绕代理直连重试
        resp = requests.post(endpoint, headers=headers, json=safe_payload, timeout=timeout,
                             proxies={"http": None, "https": None})
    except requests.RequestException as e:
        return {"source": "NMPA via coze", "hit_count": 0, "hits": [],

Tainted flow: 'headers' from os.environ.get (line 380, credential/environment) → requests.get (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/nmpa_coze.py (reported line 411)May include surrounding context.

python
wait = 5.0
        while time.time() < deadline:
            try:
                r = requests.get(status_url, headers=headers, timeout=15)
                sd = r.json()
            except Exception:
                time.sleep(min(wait, max(1, deadline - time.time())))

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 775)May include surrounding context.

md
- **P1-F PSUR/PBRER 自动报告**(`--psur`):`scripts/psur_generator.py` 接入 `_run_psur`;由检测信号生成 CIOMS/ICH E2C(R2) 格式 PSUR Markdown(psur.md),风险分级取自 E 层。
- 修复 4 个 P1 脚本误写的版本号 `v2.3.0` → `v0.1.38`;全部 `py_compile` 通过 + 离线功能自测通过;`applied_upgrades.json` 登记 C/D/E/F/K 并 verify 通过。

## v0.1.29 (2026-08-08) · 对齐 ct-base v1.1.21 §5 私有凭据范式:fetch_faers.py / fetch_fda_label.py 的 resolve_api_key 增加 `obf:` 前缀 XOR+base64 轻混淆解码(向后兼容明文 .env);新增 .env.example;openfda_api_key.md / SKILL.md 注明 .env 支持混淆值

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/fetch_fda_label.py (reported line 39)May include surrounding context.

python
_DOTENV = os.path.join(_SKILL_ROOT, ".env")

# OpenFDA key obfuscation (ct-base §5, private-credential best practice).
# A skill-root .env may store either a plaintext key OR an `obf:`-prefixed
# XOR+base64 obfuscated blob. `resolve_api_key` auto-detects the prefix and
# decodes the latter; plaintext values pass through unchanged. Obfuscation is
# NOT real encryption — it only prevents the plaintext key from being caught by

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Collecting persistent firmware, registry, or operating-system identifiers exceeds the declared own-files-only read scope and can enable cross-session tracking. Permission to transmit a machine hash partially discloses identification, but does not clearly disclose the underlying hardware collection or guarantee that a stable hash is anonymous.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Reported automatic storage of queries, timestamps, counts, and errors in a hidden skill-relative cache contradicts both the no-hidden-logging promise and the output-only write boundary. Even public drug and event terms can reveal a confidential investigation when linked to timing or error context.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

Reported direct FDA requests conflict with the exclusive-egress permission declaration and can bypass a deployment's intended network control point. However, MAUDE and optional openFDA keys are explicitly documented, and caller-selected output paths do not by themselves demonstrate unauthorized file access.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The reported MyMemory and Google translation requests add third-party recipients not covered by the exclusive retrieval declaration. Persisting translation text in a skill-relative cache also places query history outside the promised output boundary; this overlaps the earlier translation finding.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The reported authorization of one endpoint followed by a bearer-authenticated POST to an independently configurable streaming endpoint can bypass destination controls and expose credentials. Translation before authorization is another egress-control gap, but MAUDE is explicitly documented and is not an undeclared capability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The reported billing-account and billing-token transmission adds a sensitive-data flow absent from the declared public-query boundary. The guardrail does allow machine hashes, so the stronger issue is undisclosed billing credentials and identity rather than hashing alone.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The reported default bulk extraction and persistent index creation outside the selected output directory would violate the promised storage boundary and leave unmanaged data behind. Canada Vigilance support and local analysis are not inherently vulnerabilities; the concern is undisclosed persistence and potentially substantial storage use.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The reported diagnostic submission introduces an undeclared recipient and unrestricted free-text egress, creating a path for confidential clinical or organizational information to leave the machine. Machine hashes are expressly permitted in the guardrail, so identifiers alone do not establish this finding; the additional destination and uncontrolled diagnostic content are the substantive risks.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The reported direct label queries undermine the stated exclusive Coze egress boundary if deployments rely on that declaration. An optional FDA credential is explicitly supported by the document, so its legitimate use with FDA is not independently evidence of credential theft.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The reported language setters modify skill-owned state despite an explicit read-only declaration, so installations relying on that boundary receive unexpected mutations. The apparent purpose is benign preference persistence, and the evidence does not show attacker-controlled configuration changes or code execution.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The local statistical analyses plausibly support the declared safety-screening workflow; the absence of disproportionality calculations or public-source retrieval in this individual chunk is not itself evidence of a mismatch. However, its filesystem behavior contradicts explicit permission limits: open(args.data) accepts unrestricted input paths, and open(args.output, 'w') accepts unrestricted output paths, with no --out-dir mechanism or path containment check. These operations require user-supplied arguments, and there is no evidence of covert access, hidden logging, or data transmission.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

Reported translation fallbacks send query text to recipients outside the declared egress route, potentially exposing sensitive investigation terms. A persistent skill-directory cache also violates the stated write boundary and retains data outside the user's selected output location.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The reported adapter-import failure leaves alternate retrieval modules available, creating a fail-open path around the declared outbound architecture. MAUDE is explicitly documented, and caller-supplied local analysis inputs are not inherently dangerous, so those portions do not substantiate the vulnerability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The reported temporary-file lifecycle leaves generated R source behind on a rejection path and places it outside the promised output directory. This creates an avoidable retention issue, although the finding does not demonstrate code execution, weak temporary-file permissions, or disclosure to another user.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

MedDRA coding is a plausible supporting component of adverse-event signal analysis, so its narrower functionality alone is not a mismatch. However, _load_meddra_file opens a supplied filesystem path, contrary to the own-files-only read restriction. main also opens the supplied --output path in write mode without output-directory confinement, permitting creation or truncation of any file writable by the process. These are concrete filesystem-boundary discrepancies. Although the module mentions an optional LLM mode, use_llm is only stored and no API call is implemented; there is no evidence of network transmission or hidden logging.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Report generation is a plausible supporting function for safety-signal analysis, so its lack of retrieval or statistical calculations is not itself a mismatch. The filesystem behavior, however, conflicts with explicit permission limits: neither the input path nor the output path is checked against an allowed directory, and this script provides no --out-dir confinement. No network transmission, hidden logging, or autonomous file discovery is present in the supplied chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

Office conversion could be a supporting implementation detail, so the absence of FAERS analysis in this individual module is not by itself decisive. The concrete mismatch is its input and filesystem scope: the script accepts arbitrary accessible Office-file paths without restricting them to the skill’s own files or checking that their contents are public. PPTX processing additionally exposes text from ancillary and embedded parts. No transmission, hidden logging, or filesystem writes are demonstrated by this chunk.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The statistical functionality matches the declared pharmacovigilance purpose, and the supporting mapping and validation features are related implementation details. However, their filesystem access exceeds the explicitly declared boundary: input paths are not confined to the skill's files, and open(args.out, "w") can write anywhere permitted by the process rather than only within an approved output directory. No exfiltration, hidden logging, or unrelated execution is visible. The absence of retrieval and Safety Signal Score logic in this individual module is not itself evidence of a mismatch.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The qualitative causality heuristic is a plausible supporting component of safety-signal analysis; its lack of PRR/ROR/IC/EBGM calculations in this chunk is not itself a mismatch. The concrete mismatch is filesystem scope: _load_input directly opens the supplied input path, and main directly opens --json-out for writing, potentially creating or overwriting any file permitted by the process's operating-system permissions. Neither operation enforces the declared filesystem boundaries. There is no evidence of network transmission, hidden logging, or exfiltration in this chunk.

Content

No source excerpt is available for this finding.