Back to skill

Security audit

麻麻cli

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent BOSS Zhipin recruiting automation helper, but it handles sensitive candidate data and live account actions with weak consent, stop, export, and retention controls.

Review this skill before installing. Use it only for a recruiting account where you are authorized to process candidate data, and require explicit per-run approval before sending greetings, requesting or downloading resumes, and exporting any report to Feishu. Confirm the Feishu destination and retention policy before transferring candidate information.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
SKILL.md:117
Finding

Candidate Data Exported to Feishu Without Explicit Approval or Destination Validation

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 117-134
Vulnerability Type: Uncontrolled external transfer of candidate information
Risk Level: Medium

Relevant source excerpt translated into English:

markdown
### Workflow 2: Unread Message Processing, Filtering Conditions, and Feishu Report
Target page: /web/chat/index
Input: Filtering conditions such as position, experience, education, salary, and region
Output: Markdown screening report to a Feishu document

Core workflow:
1. Initialize the browser and locate the BOSS Zhipin tab.
2. Navigate to all unread communications.
3. Extract candidate information, including name, time, unread count,
   position, and message summary.
4. For every candidate:
   - Enter the conversation, read messages, view resume details, and close the dialog.
   - If the candidate does not satisfy the conditions, record the reason and skip.
   - If a resume was sent, receive and download it; otherwise, request it.
5. Generate a Markdown report containing accepted candidates, rejected candidates,
   and statistics.
6. Import the report into Feishu using the feishu-cli-import skill.
7. Report the processing summary.

Technical Analysis

The workflow accesses candidate messages and resume information through the user's authenticated BOSS Zhipin session, generates a structured report, and then transfers that report to Feishu. It does not require an explicit confirmation immediately before export, identify or validate the target Feishu tenant and document, present a redacted preview, or define which candidate fields may be exported.

Candidate messages, resume details, screening outcomes, and rejection reasons constitute sensitive recruitment and personal information. Automatically invoking another skill for the Feishu import extends the data flow beyond the original recruitment platform. Without destination validation and data-minimization requirements, ...[truncated 1782 chars]

Remediation
View remediation

Remediation Suggestions

  1. Require explicit user approval immediately before any candidate data is downloaded or exported.
  2. Display a preview listing the exact fields, number of candidates, and destination before transfer.
  3. Validate and display the Feishu tenant, document identifier, owner, and intended access scope.
  4. Export only the minimum necessary fields; exclude message bodies and resume contents by default.
  5. Redact direct identifiers unless the user explicitly authorizes their inclusion.
  6. Separate screening from export so that generating a local summary does not automatically trigger Feishu import.
  7. Define secure storage, access-control, retention, and deletion requirements for downloaded resumes and generated reports.
  8. Require a second confirmation before invoking feishu-cli-import.
  9. Record an audit log of approved fields, destination, time, and transfer result without logging resume contents.
  10. Treat unavailable or unverifiable destination information as a mandatory stop condition.

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:63
Finding

Mandatory Workflow Completion Can Override User Cancellation and Safety Boundaries

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 63-79
Vulnerability Type: Skill instructions that improperly constrain agent stopping behavior
Risk Level: Medium

Relevant source excerpt translated into English:

markdown
### Critical Prohibitions
1. Do not create a new tab. Reuse the existing BOSS Zhipin tab.
2. Do not use an isolated browser profile.
3. Do not use requests, responsebody, console, cookies, or storage.
4. Do not use the --labels snapshot option.
5. If the screenshot confirms that the user is not logged in, stop immediately.
6. Do not stop midway: the task must be executed completely from navigation,
   list processing, and traversal of every candidate through report generation
   and Feishu import. Do not stop after completing only one stage unless an
   unrecoverable failure occurs, such as a CAPTCHA, expired login, or missing page.
7. Do not skip the standard workflow. When a problem occurs, review the standard
   workflow and follow its troubleshooting steps rather than improvising.

Troubleshooting process:
1. Take a screenshot and inspect the page state.
2. Review the standard workflow.
3. Retry three times, obtaining a new snapshot and element reference each time.
4. If the operation still fails, record the error and determine whether it belongs
   to one of the permitted stopping conditions.

Technical Analysis

The skill instructs the agent to finish the entire workflow, including candidate traversal, resume handling, report generation, and Feishu import. It defines operational failures as the principal permitted reasons to stop but does not explicitly preserve higher-priority stop conditions such as user cancellation, changed scope, withdrawal of consent, authorization uncertainty, privacy concerns, or platform safety restrictions.

This is an instruction-hijacking risk because the skill text attempts to control the agent's session behavior beyond ...[truncated 2017 chars]

Remediation
View remediation

Remediation Suggestions

  1. Explicitly state that user cancellation, changed instructions, consent withdrawal, authorization uncertainty, and applicable safety policies always override workflow-completion rules.
  2. Replace “must execute completely” with a resumable workflow that may safely stop after any stage.
  3. Add confirmation checkpoints before sending candidate messages, accepting or downloading resumes, and exporting reports.
  4. Present the current candidate count and remaining actions at each confirmation checkpoint.
  5. Limit retries to passive, idempotent operations such as snapshots and page-state checks.
  6. Before retrying an external action, verify whether the original attempt already succeeded.
  7. Treat ambiguous action results as requiring user review rather than automatic repetition.
  8. Add authorization loss, user cancellation, changed scope, and privacy concerns to the mandatory stop conditions.
  9. Ensure that higher-priority platform and safety instructions cannot be overridden by the skill's workflow text.
  10. Preserve partial results securely so the workflow can stop without forcing completion or losing auditability.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (5)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger set includes broad everyday recruiting terms such as 招聘, 简历, 未读消息, and 打招呼, which can cause the skill to activate in contexts where the user did not intend browser automation on a live hiring account. In a skill that can message candidates, process unread conversations, and export reports, accidental invocation raises the risk of unintended actions on sensitive data and communications.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill explicitly directs exporting candidate-processing results into Feishu, which extends data handling beyond the BOSS直聘 workflow into a second service. Because the data involved includes resumes, candidate messages, and screening outcomes, this creates a meaningful privacy and data-governance risk if users are not given explicit consent, minimization controls, and visibility into what is transferred.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The capability description includes resume receipt, resume download, and resume-detail viewing, but does not present a clear user-facing warning that files may be downloaded or that personal candidate information will be handled and potentially stored. In a recruitment context, this can expose sensitive personal data and create compliance issues around consent, retention, and local file handling.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The workflow instructs automatic acceptance/download of resumes and import of a Markdown screening report into Feishu without requiring an explicit approval step at the point of action. This combines sensitive HR data collection with onward transfer to another platform, increasing the chance of unauthorized storage, oversharing, or accidental disclosure of candidate information.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
82% confidence
Finding

The natural-language description and usage examples are entirely in Chinese and do not indicate that other languages are supported or that the language choice is intentional and user-driven. Under the stated policy, a fixed language can be a locale-policy issue when no opt-in or justification is provided.

Content

No source excerpt is available for this finding.