Back to skill

Security audit

可转债条款速读

Security checks for vulnerabilities and agentic risk

Overview

Scanners rate this skill suspicious with medium confidence, citing an unexplained MCP discovery capability, unpinned installation commands, and privacy and validation concerns; inspection was partial because one referenced path was missing.

Consider avoiding installation until the unrelated MCP discovery behavior is explained or removed, the installer and source are pinned, and the document-processing privacy expectations and validator issue are addressed. The scan covered all listed components but was marked partial because a referenced path was missing, so the evidence is incomplete.

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
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (17)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied chunk performs MCP service discovery and connectivity checks, not convertible-bond document extraction or terms-card generation. No connection is shown between these network requests and obtaining or analyzing the specified bond documents, so they cannot be established as supporting implementation details. The authenticated tool-enumeration behavior is an undeclared capability. The chunk does not demonstrate investment advice, trading, or misuse of the credential beyond sending it to the configured endpoint.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The supplied chunk implements MCP connectivity and tool-discovery checks. It does not read a prospectus, extract the eight required term categories, or produce citation-backed term cards. Nothing in the chunk connects the external catalog or authenticated tool enumeration to the declared workflow, so this is a materially different purpose rather than an evidenced supporting implementation detail.

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 27)May include surrounding context.

3. Install and run

bash
mkdir -p ~/.workbuddy/skills
cp -r convertible-bond-terms-cn ~/.workbuddy/skills/
# ask the agent for the five-section card defined in SKILL.md, save it as my-cb-card.md
python3 ~/.workbuddy/skills/convertible-bond-terms-cn/scripts/check_cb_card.py ./my-cb-card.md

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 28)May include surrounding context.

bash
# 1. 放到技能目录
mkdir -p ~/.workbuddy/skills
cp -r convertible-bond-terms-cn ~/.workbuddy/skills/

# 2. 让 agent 按 SKILL.md 的条款卡五段出卡,产物存成我的转债卡.md

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 installs skill content from a mutable repository reference. Although this is a skills installer rather than an MCP server itself, compromised or unexpectedly changed upstream content could execute with local privileges or introduce persistent 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

The optional collection installation uses an unpinned executable and a mutable upstream skill collection. Optional setup reduces exposure, but users who follow it can receive unreviewed executable code or 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 a single skill limits installation scope but does not pin either the installer or the source revision. An upstream compromise could therefore substitute malicious code or instructions during setup.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · README.en.md (reported line 55)May include surrounding context.

2) take routing.url and transport for the domain from the response; never hard-code the host

3) with the key, send tools/list - proves that endpoint accepted the key, nothing else

curl -s -X POST "$ROUTING_URL" -H "Authorization: Bearer $CUE_API_KEY"
-H "Content-Type: application/json" -H "Accept: application/json, text/event-stream"
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | head -c 400

text

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · README.md (reported line 59)May include surrounding context.

2) 域路由现值:从返回里取该域的 routing.url 与 transport,别把域名抄死

3) 带 key 实连后发 tools/list——只证该数据腿端点被接受,不证别的入口

curl -s -X POST "$ROUTING_URL" -H "Authorization: Bearer $CUE_API_KEY"
-H "Content-Type: application/json" -H "Accept: application/json, text/event-stream"
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | head -c 400

text

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

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The installation command resolves an unpinned npm executable and mutable skill repository content. This creates supply-chain exposure to local code execution through the installer and unreviewed instructions through the installed skill.

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 collection installation does not fix the installer version or upstream revision. Even though users can select individual skills, compromised upstream material could still introduce executable code or persistent 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

The single-skill selector narrows scope but leaves both executable and skill content mutable. Upstream compromise or an unexpected release could change what runs locally or what instructions the agent subsequently loads.

Content

No source excerpt is available for this finding.

The attachment-parsing recovery step lacks a warning about sending supplied documents to the third-party Cue service.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The recovery instruction directs attachment parsing without requiring disclosure or consent before supplied content goes to an external service. Prospectuses are usually public, which reduces typical exposure, but user attachments can contain annotations or other nonpublic information; a local no-storage statement does not address third-party processing.

Content

No source excerpt is available for this finding.

The document-processing workflow lacks a privacy warning about sending supplied documents to external services.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The workflow sends long attachments through an external parsing channel without a pre-upload privacy notice or authorization requirement. Public bond disclosures lower the usual sensitivity, but the accepted user-supplied documents can include nonpublic content that would cross an external trust boundary.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write.

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 attachment-parsing workflow lacks a privacy warning about sending document content to an external service.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The skill initiates external attachment parsing when needed but does not require informed authorization before transmitting supplied content. Limiting parsing to relevant attachments reduces volume, not confidentiality risk, and user-supplied material may contain private annotations despite the task's focus on public disclosures.

Content

No source excerpt is available for this finding.

The documented line-prefix requirement for [待人工] is not enforced, allowing embedded markers to bypass financial-advice checks.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

PENDING.search(ln) exempts an entire line from CB-LINE checks whenever [待人工] appears anywhere, despite the comment specifying a line-prefix marker. An input author can append the marker to a prohibited recommendation or guaranteed-return claim and suppress those checks, potentially allowing the card to pass if its other requirements are satisfied. This undermines a central financial-content safeguard in this skill, although it does not enable code execution or financial transactions.

Content

No source excerpt is available for this finding.