ClawHub
Back to skill

Security audit

Humanizer (Polish)

Security checks for vulnerabilities and agentic risk

Overview

This Polish text-humanizing skill is mostly coherent, but its editing instructions can cause an agent to add unsupported facts or author opinions directly into documents.

Install only if you are comfortable reviewing every proposed rewrite before saving it. Use it on clearly selected passages, and do not rely on it for legal, academic, technical, compliance, or public-facing text unless you verify that no facts, sources, opinions, or first-person claims were added.

A.I.G

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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)

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:43
Finding

Unsupported Fact and Persona Injection During Text Rewriting

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:43-61, with unsafe examples at SKILL.md:110, SKILL.md:138, and SKILL.md:421; the last example is repeated at README.md:73
Vulnerability Type: Content-integrity failure caused by unsafe rewriting instructions
Risk Level: Medium

Complete Vulnerable Snippets

The skill directs the agent to add opinions, first-person statements, emotions, and intentionally imperfect material:

text
**Miej zdanie.** Nie relacjonuj tylko faktów — reaguj na nie. „Szczerze nie wiem, co o tym myśleć" jest bardziej ludzkie niż neutralne wylistowanie za i przeciw.

**Różnicuj rytm.** Krótkie zdania. Potem dłuższe, które nie spieszą się, żeby dojść do sedna. Mieszaj.

**Przyznawaj się do złożoności.** Ludzie mają mieszane uczucia. „To robi wrażenie, ale trochę niepokoi" wygrywa z „To robi wrażenie."

**Używaj „ja", gdy to pasuje.** Pierwsza osoba nie jest nieprofesjonalna — jest uczciwa. „Ciągle wracam do…" albo „Nie daje mi spokoju…" sygnalizuje, że za tekstem stoi żywy człowiek.

**Pozwól na odrobinę bałaganu.** Idealna struktura wygląda algorytmicznie. Dygresje, wtrącenia i nie do końca uformowane myśli — to jest ludzkie.

**Bądź konkretny w uczuciach.** Nie „to budzi obawy", ale „jest coś niepokojącego w agentach, które mielą kod o trzeciej w nocy, kiedy nikt nie patrzy."

Several recommended rewrites introduce unsupported facts or attributions:

text
Świątynia jest utrzymana w kolorach niebieskim, zielonym i złotym. Architekt powiedział, że nawiązują one do lokalnych chabrów i wybrzeża Zatoki.
text
Rzeka Haolai jest siedliskiem kilku endemicznych gatunków ryb, według badania Chińskiej Akademii Nauk z 2019 roku.
text
Aktualizacja dodaje przetwarzanie wsadowe, skróty klawiaturowe i tryb offline. Wstępne opinie beta testerów są pozytywne — większość zgłasza szybsze wykonywanie zadań.

Technical Analysis

The skill states that rewr ...[truncated 2665 chars]

Remediation
View remediation

Remediation Suggestions

  1. Add an explicit invariant that the agent must not introduce facts, sources, quotations, metrics, product features, opinions, emotions, or first-person claims absent from the input.
  2. Replace every example that invents attribution or factual detail with a rewrite that preserves all original factual assertions.
  3. Instruct the agent to retain vague wording when the source is vague, or to ask the user for the missing details rather than generating them.
  4. Require placeholders such as [source needed] or [confirm specific feature] when concrete information is necessary but unavailable.
  5. Distinguish stylistic editing from substantive editing. Any semantic addition should require explicit user approval.
  6. Before using Write or Edit, compare the proposed result with the source and flag newly introduced names, dates, citations, quotations, numbers, capabilities, and first-person assertions.
  7. Default to returning a proposed rewrite for review rather than modifying the original file when the factual-difference check detects semantic changes.
  8. Add regression tests containing factual, legal, academic, and technical passages to verify that humanization changes style without adding unsupported propositions.

SkillSpector

Vulnerability Patterns
  • 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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (2)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The README says the skill can be invoked 'bezpośrednio podczas edycji dokumentów,' which is a broad trigger lacking boundaries, scope, or confirmation requirements. In an agentic environment, that can cause the skill to run on ordinary editing activity and rewrite text unexpectedly, creating integrity and workflow risks even if the skill itself is not overtly malicious.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
77% confidence
Finding

The example trigger 'Zhumanizuj ten tekst: [twój tekst]' is understandable, but it does not define what content is in scope, what should not be transformed, or when the skill should refuse. That ambiguity can lead to over-application on sensitive, formal, or citation-bound text where stylistic rewriting may alter meaning or provenance.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.