Back to skill

Security audit

论文研读卡·观点可追句、引文可回源

Security checks for vulnerabilities and agentic risk

Overview

The paper-reading workflow is coherent, but its optional connection setup can expose your API key to an unintended destination.

Use the local/manual workflow without CUE_API_KEY until both connection examples enforce trusted HTTPS destinations and authenticated redirects are blocked or safely revalidated. Before external parsing, confirm what document content will be sent and how the provider handles it, especially for confidential manuscripts. Review exact installer and repository revisions and install only the needed parser skill. The separately installed parser and live service were not bundled or assessed.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Rogue AgentSelf-Modification, Session Persistence
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
Findings (27)

Tainted flow: 'req' from os.environ.get (line 188, 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 · scripts/cue_connectivity_check.py (reported line 193)May include surrounding context.

python
code = None
    raw = ""
    try:
        with urllib.request.urlopen(req, timeout=args.timeout) as resp:
            code, raw = resp.status, resp.read().decode("utf-8", "replace")
    except urllib.error.HTTPError as err:
        detail = ""

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · README.en.md (reported line 88)May include surrounding context.

bash
   CATALOG="https://cuecue.cn/api/mcp-catalog"
   curl -s "$CATALOG" | python3 -c 'import json,sys
   rows=json.load(sys.stdin)["data"]["domains"]
   [print(r.get("id"), r.get("external_status"), (r.get("routing") or {}).get("url", "-")) for r in rows]'
   DOMAIN_ID="<replace with one of the ids listed by the line above>"

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · README.en.md (reported line 92)May include surrounding context.

md
rows=json.load(sys.stdin)["data"]["domains"]
   [print(r.get("id"), r.get("external_status"), (r.get("routing") or {}).get("url", "-")) for r in rows]'
   DOMAIN_ID="<replace with one of the ids listed by the line above>"
   ROUTING_URL="$(curl -s "$CATALOG" | python3 -c 'import json,sys
   rows=json.load(sys.stdin)["data"]["domains"]
   hit=[r["routing"]["url"] for r in rows if r.get("id")==sys.argv[1] and r.get("routing")]
   print(*hit,sep="\n") if hit else sys.exit("Not found: that id is not in the catalog response (or its routing has no url field) - run the line above and pick one of the ids listed")' "$DOMAIN_ID")"

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · README.md (reported line 86)May include surrounding context.

bash
   CATALOG="https://cuecue.cn/api/mcp-catalog"
   curl -s "$CATALOG" | python3 -c 'import json,sys
   rows=json.load(sys.stdin)["data"]["domains"]
   [print(r.get("id"), r.get("external_status"), (r.get("routing") or {}).get("url", "-")) for r in rows]'
   DOMAIN_ID="<把这里换成上一行列出的那个域 id>"

External Script Fetching

High
Category
Supply Chain
Confidence
90% confidence
Finding

Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.

Content

Scanner excerpt · README.md (reported line 90)May include surrounding context.

md
rows=json.load(sys.stdin)["data"]["domains"]
   [print(r.get("id"), r.get("external_status"), (r.get("routing") or {}).get("url", "-")) for r in rows]'
   DOMAIN_ID="<把这里换成上一行列出的那个域 id>"
   ROUTING_URL="$(curl -s "$CATALOG" | python3 -c 'import json,sys
   rows=json.load(sys.stdin)["data"]["domains"]
   hit=[r["routing"]["url"] for r in rows if r.get("id")==sys.argv[1] and r.get("routing")]
   print(*hit,sep="\n") if hit else sys.exit("未取到:该 id 不在目录现值里(或它没有 routing 里的 url 字段)——先跑上面那行列出现场的 id 再填")' "$DOMAIN_ID")"

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied code has a materially different primary purpose from the declared academic-reading workflow. It neither processes papers nor produces any of the promised reading outputs; it discovers and tests MCP service endpoints. Although authenticated POST requests require --post and an HTTPS/host-suffix check, those safeguards do not make credential access and service probing part of the declared purpose. The default path also performs an anonymous catalog GET unless a local catalog file is supplied, despite comments claiming no packets are sent by default. No connection between this diagnostic utility and the declared paper-reading functionality is established in the supplied material.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied shell snippet performs generic MCP service discovery and authenticated tool enumeration, not academic-paper analysis. It does not process a PDF, extract page references, check citations, or generate reading materials. Although tool discovery could support a larger workflow, this chunk establishes no connection between the selected service and the declared paper-reading purpose. Its external catalog access and credential-bearing request therefore represent undeclared behavior. The snippet alone does not establish malicious intent or credential theft.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The supplied code implements generic authenticated MCP discovery rather than the declared paper-reading workflow. It does not retrieve or analyze a paper, create page-linked notes, inspect reproducibility information, or verify citations. Although service discovery could support a larger implementation, this chunk does not establish any paper-specific relationship or constrain the selected service accordingly. The domain placeholder must be replaced before the authenticated request can succeed; this does not change the capability represented by the code.

Content

No source excerpt is available for this finding.

Automatic redirects can bypass the documented HTTPS and authorized-host guarantees for credential transmission.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

The destination guard validates only the initial URL, while urllib.request.urlopen uses automatic redirect handling. For supported redirects such as HTTP 302 or 303, Python can convert the POST to a GET while retaining its Authorization header, allowing an approved endpoint to redirect the Bearer token to an unapproved host or an HTTP URL. Exploitation requires control of an approved endpoint or a usable redirect there, but can disclose the service credential despite the documented safeguards.

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.

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · README.en.md (reported line 44)May include surrounding context.

bash
# 1. Put it in the skills directory
mkdir -p ~/.workbuddy/skills
cp -r paper-study-cn ~/.workbuddy/skills/

# 2. Copy the card template and fill it following section 1

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · README.md (reported line 42)May include surrounding context.

bash
# 1. 放到技能目录
mkdir -p ~/.workbuddy/skills
cp -r paper-study-cn ~/.workbuddy/skills/

# 2. 复制卡模板,按 §一 三步填

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

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The installation command invokes an unpinned npm CLI and retrieves a mutable skill collection. Although this is a skill installer rather than an MCP server reference, a compromised future release could execute code during installation or introduce adversarial agent instructions.

Content

No source excerpt is available for this finding.

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

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

Selecting only cue-omni-reader reduces installation scope but does not pin either the npm installer or the retrieved skill revision. A compromised upstream update could therefore change executable installation behavior or the instructions subsequently trusted by the agent.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
98% confidence
Finding

The shell example sends CUE_API_KEY to a URL taken directly from a remote catalog without validating its scheme or authorized host. A compromised or malicious catalog response could redirect the bearer credential to an attacker or an HTTP endpoint; the separately described Python helper's destination checks do not protect this curl example.

Content

Scanner excerpt · README.en.md (reported line 97)May include surrounding context.

md
hit=[r["routing"]["url"] for r in rows if r.get("id")==sys.argv[1] and r.get("routing")]
   print(*hit,sep="\n") if hit else sys.exit("Not found: that id is not in the catalog response (or its routing has no url field) - run the line above and pick one of the ids listed")' "$DOMAIN_ID")"
   [ -n "$ROUTING_URL" ] || { echo "Fetch failed: ROUTING_URL is empty - run the line above, then set DOMAIN_ID to one of the ids listed" >&2; exit 1; }
   curl -s -X POST "$ROUTING_URL" \
     -H "Authorization: Bearer $CUE_API_KEY" \
     -H "Accept: application/json, text/event-stream" \
     -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' \

The parsing workflow requests spending consent but does not warn that paper content is sent to an external service.

Medium
Category
Output Handling
Confidence
92% confidence
Finding

The workflow requests consent for parsing costs but does not explicitly request consent to disclose document contents to the external parsing provider. Since local files and pasted text are accepted, users could unintentionally disclose unpublished or confidential research; the optional, manual path reduces but does not eliminate this risk.

Content

No source excerpt is available for this finding.

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

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The package installation example uses an unpinned npx installer and a mutable repository reference. This creates a supply-chain exposure even though the command installs only the intended paper-study skill.

Content

No source excerpt is available for this finding.

The optional Cue parsing workflow lacks a warning that user-provided document content will be sent to a third-party service.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The optional external parsing workflow accepts local files or pasted text but does not explicitly explain that document contents may leave the user's environment for a third-party service. Confirmation of parsing costs is not informed consent to disclose potentially confidential manuscripts or embedded personal information.

Content

No source excerpt is available for this finding.

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

Medium
Category
MCP Rug Pull
Confidence
96% confidence
Finding

The installation command invokes the npm-distributed skills CLI without a version pin and installs from a mutable repository reference. Although this is not an MCP server declaration, a compromised or unexpected upstream release could execute installer code locally or install adversarial agent instructions.

Content

No source excerpt is available for this finding.

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

Medium
Category
MCP Rug Pull
Confidence
96% confidence
Finding

Selecting only cue-omni-reader limits installation scope but does not pin either the executed npm installer or the upstream skill contents. Upstream compromise or release drift could therefore introduce local executable code or malicious agent instructions.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
99% confidence
Finding

The copyable shell workflow sends CUE_API_KEY to a URL obtained from the remote catalog without checking its scheme or destination host. A compromised catalog could direct the bearer credential to an attacker or an unencrypted HTTP endpoint; the separately described Python helper's safeguards do not protect this curl command.

Content

Scanner excerpt · README.md (reported line 95)May include surrounding context.

md
hit=[r["routing"]["url"] for r in rows if r.get("id")==sys.argv[1] and r.get("routing")]
   print(*hit,sep="\n") if hit else sys.exit("未取到:该 id 不在目录现值里(或它没有 routing 里的 url 字段)——先跑上面那行列出现场的 id 再填")' "$DOMAIN_ID")"
   [ -n "$ROUTING_URL" ] || { echo "取值失败:ROUTING_URL 为空——先跑上面那行列出现场 id,再把 DOMAIN_ID 换成其中之一" >&2; exit 1; }
   curl -s -X POST "$ROUTING_URL" \
     -H "Authorization: Bearer $CUE_API_KEY" \
     -H "Accept: application/json, text/event-stream" \
     -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' \

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

Medium
Category
MCP Rug Pull
Confidence
96% confidence
Finding

The self-installation command uses an unversioned npm CLI and a mutable source repository reference. Checking that the skill name appears afterward establishes installation presence, not the integrity of executable installer code or installed agent instructions.

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 workflow routes papers through a parsing service without warning that document contents may be disclosed to an external provider.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

The input contract accepts local files and pasted text, while the parsing workflow does not explicitly disclose that their contents may leave the user's environment or require informed upload consent. Mentioning a parsing channel and configuration page does not adequately protect confidential manuscripts or research data from unintended external disclosure.

Content

No source excerpt is available for this finding.

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_study_card.py (reported line 333)May include surrounding context.

python
scripts_dir.mkdir(parents=True)
        dst = scripts_dir / Path(__file__).name
        shutil.copy2(str(Path(__file__).resolve()), str(dst))
        r = subprocess.run([sys.executable, '-B', str(dst), str(card)], capture_output=True, text=True)
        out = r.stdout + r.stderr
        named = '尺子的来源不可读' in out
        detail = '隔离副本(无 assets/ 模板)退码 %s,详行含「尺子的来源不可读」=%s' % (