Back to skill

Security audit

去除文本中

Security checks for vulnerabilities and agentic risk

Overview

This skill is a text humanizer that openly focuses on removing AI-writing traces, but it asks for command and file-write authority that is not needed for that purpose.

Review before installing. Use this only where transparent editing is acceptable, not to hide authorship or bypass detection rules. Prefer a version that removes exec/write permissions, narrows activation to explicit rewriting requests, and states clear acceptable-use limits.

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:34
Finding

Unnecessary Command Execution and File-Write Permissions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:34-37
Vulnerability Type: Excessive tool permissions and violation of least privilege
Risk Level: Medium

Vulnerable Code

yaml
tools:
- read
- exec
- write

The excessive permissions are reinforced by the execution-capable classification at SKILL.md:118-119:

markdown
### Availability Classification
- **Category**: MD+execute()

Technical Analysis

The skill's stated purpose is to transform supplied text using an Agent-provided language model. The reviewed file does not define an implementation, workflow, or legitimate operational requirement for running system commands or modifying local files.

Nevertheless, the skill requests both exec and write. If the hosting Agent grants the declared tools, these permissions extend the skill's authority beyond what is necessary for text transformation:

  • exec can permit local command execution with the privileges of the Agent process.
  • write can permit creation or modification of files accessible to the Agent process.
  • Their combination can permit writing a script or configuration file and subsequently executing or activating it.

This violates least-privilege principles and unnecessarily increases the impact of prompt injection, malicious user-supplied content, or unsafe future modifications to the skill. The reviewed skill does not itself contain instructions that actively exploit these permissions, so this finding concerns the exposed privilege boundary rather than a confirmed malicious payload.

Attack Path

  1. A platform loads the skill and grants the tools declared in its metadata.
  2. The skill processes attacker-controlled or otherwise untrusted text.
  3. Malicious text attempts to induce the Agent to invoke the unnecessarily available write or exec tool.
  4. If the Agent fails to isolate data from instructions or lacks command-level authorization controls, it ...[truncated 1101 chars]
Remediation
View remediation

Remediation Suggestions

  1. Remove exec and write from the declared tools. Retain only the minimum capability required to receive text and return transformed text.
  2. Remove the MD+execute() classification unless executable functionality is implemented and demonstrably required.
  3. If file export is later added, expose a narrowly scoped write operation rather than unrestricted filesystem access:
    • Restrict writes to a dedicated output directory.
    • Reject absolute paths and path traversal sequences.
    • Prevent overwriting existing files unless explicitly authorized.
    • Apply restrictive file permissions.
  4. If command execution later becomes essential:
    • Use a fixed allowlist of commands and arguments.
    • Never construct shell commands from user-controlled text.
    • Avoid invoking a shell where direct process APIs are available.
    • Run commands in a sandbox with no unnecessary network access or credentials.
    • Require explicit user confirmation before execution.
  5. Treat all transformed content as untrusted data and prevent it from being interpreted as Agent instructions.
  6. Add automated policy checks that reject skill manifests requesting tools unsupported by documented functionality.
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 (6)

Ssd 2

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The core advertised function is to remove signs of AI-generated writing and make text look human-written, which is an obfuscation capability aimed at concealing provenance. In context, this is more dangerous because the skill is packaged for broad agent use with read/write/exec tools, normalizing deceptive rewriting that could be used to evade academic, platform, compliance, or moderation checks.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill description advertises extremely broad use cases like general productivity, automation, workflow optimization, and AI orchestration, making it eligible to activate in many unrelated contexts. Overbroad trigger language increases the chance the agent will invoke this skill unexpectedly and apply rewriting behavior where the user did not explicitly request AI-obfuscating transformation.

Content

No source excerpt is available for this finding.

Ssd 2

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The paid-feature table explicitly monetizes the ability to remove AI-generation traces, reinforcing that concealment is a product goal rather than an incidental side effect. Commercializing this capability increases likelihood of intentional misuse at scale, including deceptive content laundering and circumvention of detection or attribution systems.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The usage and request sections describe inputs and outputs but do not define a precise activation boundary, so the skill could be selected for a wide range of generic content tasks. In an agent ecosystem, ambiguous scope can cause unintentional invocation of a capability designed to make text appear more human-authored, which raises misuse and policy-evasion risk.

Content

No source excerpt is available for this finding.

Ssd 2

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The FAQ claims high-accuracy identification and removal of AI text traces, which further instructs users that the skill is intended to defeat provenance detection. This makes the deceptive use case explicit and lowers barriers for misuse by presenting evasion as a normal, reliable workflow outcome.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
90% confidence
Finding

The FAQ says the skill '主要支持中文文本的处理', which imposes a language constraint in natural-language documentation. The file does not present this as an explicit user opt-in choice or justify it as a region-specific requirement.

Content

No source excerpt is available for this finding.