Back to skill

Security audit

auto-monitor

Security checks for vulnerabilities and agentic risk

Overview

This Markdown-only monitoring skill is not malware, but it asks an agent to proactively inspect systems and perform automatic repairs without clear consent, scope, or safeguards.

Review before installing. Use this only in an environment where monitoring targets and report recipients are explicitly controlled, and do not grant command execution or scheduling authority unless the skill is revised to require per-action approval, define allowed repair scripts, and provide rollback and audit safeguards.

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 (2)

other

Note
Location
SKILL.md:29
Finding

Unrestricted Recurring System and Network Reconnaissance

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 29-41
Vulnerability Type: Excessive environment reconnaissance
Risk Level: Low

Code Snippet

The following is an English translation of the relevant source instructions:

text
1. Periodic checks (every N minutes)
2. Check:
   - Disk utilization
   - Memory utilization
   - CPU load
   - Network status
   - Process status
3. Detect anomalies
   - Threshold exceeded? → Proactively alert
4. Report
   - Normal: concise report
   - Abnormal: detailed explanation and recommendations

Technical Analysis

The skill directs an agent to perform recurring inspection of network and process state without defining an authorized target, collection boundaries, redaction requirements, or retention policy. Process details may disclose service names, command-line arguments, software versions, security tools, or sensitive file paths. Network inspection may expose listening services, active connections, internal addresses, and deployment topology.

The declared tool list contains only read, so the package does not itself implement network or process enumeration. Exploitation therefore depends on the hosting agent granting additional system-observation or command-execution capabilities. The issue is an overbroad instruction rather than embedded executable malware.

Attack Path

  1. A user or automated trigger activates the monitoring skill.
  2. The agent interprets the periodic monitoring workflow as authorization to inspect network and process state.
  3. If the host exposes suitable tools, the agent enumerates processes, services, connections, or related operating-system information.
  4. Sensitive operational details enter the agent context.
  5. The details are reproduced in health reports or anomaly alerts and may be exposed to unintended recipients.

Impact Assessment

The issue does not independently grant additional operating-syst ...[truncated 391 chars]

Remediation
View remediation

Remediation Suggestions

  • Require explicit user authorization before inspecting process or network state.
  • Define an allowlist of hosts, processes, interfaces, and metrics that may be monitored.
  • Collect aggregate health metrics instead of full process arguments or connection details.
  • Redact credentials, tokens, query parameters, internal addresses, usernames, and sensitive paths from reports.
  • Define report recipients, retention periods, and access controls.
  • Prohibit inspection of unrelated users, containers, namespaces, or hosts.
  • Align the documented workflow with the declared read-only tool capability.

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:48
Finding

Unbounded Automatic Remediation Without User Approval

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 48-52 and 88-89
Vulnerability Type: Unauthorized state-changing operations and least-privilege violation
Risk Level: Medium

Code Snippet

The following is an English translation of the relevant source instructions:

markdown
## Proactivity

* Do not wait for the user to ask how the server is doing.
* Proactively report discovered problems.
* Immediately handle anything that can be automatically repaired.
markdown
4. **Integrated automated remediation**: For some simple anomalies,
   Auto Monitor can automatically execute predefined remediation scripts
   without human intervention.

Technical Analysis

The skill authorizes immediate remediation without defining which scripts are trusted, which resources may be modified, what privileges may be used, or which changes require confirmation. It also lacks dry-run behavior, rollback controls, integrity verification, execution timeouts, concurrency controls, and an audit policy.

If a host agent has command-execution or administrative capabilities, an ambiguous anomaly may cause it to choose destructive actions such as terminating processes, restarting services, changing configuration, deleting files, disabling controls, or modifying firewall rules. The reference to “predefined” scripts does not identify or include any reviewed scripts in the project.

The skill metadata grants only the read tool, and the audited project contains no executable scripts. Consequently, this package does not itself obtain elevated privileges or execute remediation. The risk materializes when a hosting platform grants broader tools than the manifest declares or treats the Markdown instructions as authorization for privileged actions.

Attack Path

  1. The skill observes or incorrectly infers a threshold violation.
  2. Its instructions classify the condition as automatically repairable.
  3. The agen ...[truncated 1109 chars]
Remediation
View remediation

Remediation Suggestions

  • Disable automatic remediation by default and require explicit approval for every state-changing action.
  • Define a narrowly scoped allowlist of reviewed remediation operations.
  • Run remediation through a dedicated least-privileged service account.
  • Require dry-run output that identifies the exact command, target, expected effect, and rollback procedure.
  • Cryptographically verify approved remediation scripts and prohibit dynamically generated shell commands.
  • Validate all parameters and avoid passing untrusted monitoring data to a shell interpreter.
  • Add execution timeouts, locking, rate limits, post-action health checks, and transactional rollback.
  • Record the initiating alert, selected action, authorization, command, result, and affected resources in an immutable audit log.
  • Explicitly prohibit destructive actions such as file deletion, security-control modification, and broad process termination unless separately authorized.
  • Remove claims of automatic script execution unless the tool manifest and implementation provide a reviewed, constrained execution mechanism.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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 (8)

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill is described as proactively monitoring and reporting 'without waiting for instructions,' with no clear activation boundaries, cadence controls, or consent model. In an agent environment, self-starting behavior can bypass normal user intent checks and lead to unauthorized observation, noisy alerts, or escalation into unsafe actions if paired with stronger tools later.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger keywords are extremely broad, including common words like 'system', 'monitor', and 'auto', which can cause unintended activation during ordinary conversation. Combined with a monitoring-themed skill, this broad matching expands the chance of the skill engaging without informed user intent, potentially exposing system context or prompting unsafe follow-on actions.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

L053 states '可以自动修复的立刻处理' ('immediately handle what can be auto-repaired'), which implies active remediation. However, the manifest only declares the 'read' tool at L009-L010, and the rest of the document describes monitoring/reporting rather than any concrete repair actions, creating a direct contradiction between stated intent and actual documented capability.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill advertises that it can automatically repair issues but provides no user-facing warning about system-altering consequences. Even if current tooling is read-only, this language normalizes autonomous remediation and can cause operators or downstream agents to assume silent modifications are acceptable, weakening safety expectations around production systems.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The document states the skill is 'MD+execute' and needs command execution even though the manifest declares only read access. This capability mismatch can mislead an agent or user into attempting privileged system actions outside the declared trust boundary, undermining least-privilege expectations and safe review of the skill.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

Advertising execution of predefined repair scripts without prominent safety warnings is risky because repair scripts can change configuration, restart services, or delete data. In a system-monitoring context, users may trust the skill to act on production infrastructure, so omission of warnings increases the chance of unsafe or unauthorized system changes.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
82% confidence
Finding

The trigger list includes both Chinese and English terms, but the document does not explain whether the skill is intended to operate bilingually or how language selection is determined. This can create a locale-policy concern because the behavior appears to impose language handling implicitly rather than through documented user choice.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
82% confidence
Finding

The '差异化优势分析' section claims predictive machine learning, visualization, and automated repair capabilities. In this skill file, there is no corresponding implementation detail or declared capability beyond 'read', so the documentation materially overstates what the skill actually does.

Content

No source excerpt is available for this finding.