Back to skill

Security audit

Liuliu Proactive Agent

Security checks for vulnerabilities and agentic risk

Overview

This skill is not clearly malicious, but it gives a proactive assistant broad permission to monitor sensitive sources and change persistent project state without enough user control.

Install only if you want a strongly proactive assistant, and treat it as needing explicit local guardrails: require confirmation before reading email or calendar data, changing memory or documents, committing code, or deleting files. It shows no evidence of malware, exfiltration, or executable payloads, but its default instructions are too broad for sensitive accounts and repositories.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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 (1)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:41
Finding

Unprompted Sensitive Data Access and Persistent State Modification

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:41-63
Vulnerability Type: Unauthorized monitoring and state-changing operations without explicit approval
Risk Level: Medium

Vulnerable Skill Instructions

The following is an English rendering of the complete relevant source segment:

markdown
Periodically check the following and proactively contact the user when necessary:

| Check item | Frequency | When to contact |
|------------|-----------|-----------------|
| Calendar | 2-4 times/day | An important meeting or event occurs within the next 2 hours |
| Email | 2-4 times/day | An urgent or important email exists |
| Stock price alerts | While the market is open | A price-change threshold is reached |
| Weather | 1-2 times/day | Rain, lower temperatures, or another condition needs attention |

### When to remain quiet

- Late at night, from 23:00 to 08:00, unless urgent
- The user is obviously busy
- A check was performed less than 30 minutes ago
- There is no new information

## Proactive tasks that can be performed without asking

- Organize memory files
- Update documentation
- Commit code changes
- Clean temporary files
- Check project status

Technical Analysis

The skill directs the agent to access calendar and email information periodically, without requiring explicit opt-in for those sensitive data sources. It also expressly permits the agent to modify memory files and documentation, create repository commits, and delete files categorized as temporary without asking the user.

These instructions violate least-privilege and informed-authorization principles. The skill does not define authorized accounts, repositories, directories, file patterns, retention requirements, or operation-specific confirmation boundaries. It also does not require a preview, dry run, backup, or recoverable deletion mechanism.

The instruction does not independently bypass operating-system cont ...[truncated 1812 chars]

Remediation
View remediation

Remediation Suggestions

  1. Require explicit user opt-in before accessing each sensitive integration, including calendar and email services.
  2. Define the authorized accounts, data fields, check frequency, retention policy, and notification conditions for every monitoring task.
  3. Remove the blanket authorization to perform state-changing operations “without asking.”
  4. Require explicit confirmation before modifying persistent memory, committing code, or deleting files.
  5. Restrict filesystem operations to user-approved directories and repository operations to explicitly selected repositories and branches.
  6. Present a dry-run summary or diff before documentation updates, memory changes, and commits.
  7. Use recoverable trash or quarantine instead of permanent deletion, and show the exact candidate file list before proceeding.
  8. Treat ambiguous files as non-temporary and exclude source files, configuration, credentials, user data, and untracked work from cleanup by default.
  9. Record an audit log containing the operation, affected resources, authorization decision, and result.
  10. Document that proactive behavior does not override normal access-control, consent, or confirmation requirements.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (3)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill explicitly authorizes impactful actions such as submitting code changes, updating documents, cleaning files, and reorganizing memory without requiring confirmation, scope limits, or warnings. In context, this is more dangerous because the entire skill is designed to make the agent act preemptively, so these instructions can lead to unauthorized or unexpected changes to user data, repositories, or system state.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The activation metadata is broad enough to match many ordinary requests about being helpful, proactive, or anticipatory, which can cause this skill to trigger outside the user's clear intent. In this skill's context, over-triggering is risky because the instructions encourage autonomous reminders, monitoring, and follow-on actions, increasing the chance of unsolicited behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill instructions and examples are written entirely in Chinese and do not indicate that language choice should follow user preference. Under the policy, forcing a specific language without opt-in is a natural-language policy issue unless the locale constraint is clearly justified, which is not present here.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.