Back to skill

Security audit

Humaniseur Fr

Security checks for vulnerabilities and agentic risk

Overview

This is a French text-humanizing skill with no code execution or persistence, but it can push agents to add unsupported facts, opinions, or first-person experiences during rewrites.

Install only if you want a French-specific rewriting assistant and review outputs carefully before publication. For academic, legal, journalistic, business, or factual text, check that the rewrite has not invented statistics, citations, events, future plans, opinions, or first-person experiences.

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

other

Warning
Location
evals/evals.json:19
Finding

Unsupported Facts and Personal Experiences May Be Introduced During Rewriting

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:280, SKILL.md:303-319, SKILL.md:332; evals/evals.json:19, evals/evals.json:42, evals/evals.json:49
Vulnerability Type: Content integrity and hallucination inducement
Risk Level: Medium

Vulnerable Code Snippets

SKILL.md:280:

markdown
Replace with a concrete fact about what actually happens next.

SKILL.md:303-319:

markdown
**Avoir des opinions.** « Franchement, je ne sais pas quoi en penser » is more human than neutral pros-and-cons.

**Varier le rythme.** Short sentences that hit hard. Then longer ones with nested subordinates that take their time. French has a tradition of rhythmic asymmetry (Montaigne, Cioran, Debord). AI text is monotonously regular by contrast.

**Reconnaître la complexité.** « C'est impressionnant mais aussi un peu flippant » beats « C'est impressionnant. »

**Utiliser « je ».** First person is not unprofessional. « J'y reviens sans arrêt... » signals a human thinking. Personal voice is among the strongest authenticity markers.

**Laisser du désordre.** Perfect structure feels algorithmic. Tangents, parentheses, half-formed thoughts are human. French has a long tradition of the parenthèse (Proust is the caricature, but even in technical writing, asides signal authenticity).

**Utiliser le second degré.** LLMs are constitutionally incapable of authentic irony. Understatement, light sarcasm, self-deprecation: unfakeable markers. « On a quand même inventé un truc qui code mieux que nous quand on est fatigué, ce qui est à peu près tout le temps » does not come from an LLM.

**Être précis sur les ressentis.** Not « cela est préoccupant » but « il y a quelque chose de dérangeant à voir des agents tourner à 3h du matin sans personne pour les surveiller. »

SKILL.md:332:

markdown
- Uses specific details over vague claims

evals/evals.json:19:

json
{ "id": "1.5", "description": "rewrite 
...[truncated 2549 chars]
Remediation
View remediation

Remediation Suggestions

  1. Replace the instruction to add concrete facts with a strict source-grounding rule:
    • Preserve only facts present in the source or explicitly supplied by the user.
    • Never invent statistics, dates, quotations, citations, events, organizations, or future plans.
  2. If a vague claim cannot be made concrete from the input, either remove it, retain appropriately qualified wording, insert an explicit placeholder such as [supporting figure needed], or ask the user for evidence.
  3. Prohibit fabricated first-person experience. First-person language should be retained only when it originates in the source or is explicitly requested and supplied by the user.
  4. Clearly label hypothetical examples and suggested opinions instead of incorporating them as factual prose.
  5. Change evaluation assertion 1.5 so that it accepts only source-grounded details or explicitly labeled placeholders; it must not require information absent from the original.
  6. Revise evaluations 3.5 and related expectations to permit deletion of vague optimism without requiring an invented fact or next step.
  7. Add regression tests that fail when a rewrite introduces unsupported numbers, dates, sources, quotations, experiences, or factual claims.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The trigger list is broad enough to match ordinary editing requests such as rewriting text to sound more natural or human. That can cause the skill to activate outside the user's intended scope and silently steer responses toward this skill's rewriting style, creating prompt-routing and instruction-overreach risk rather than a classic code-execution issue.

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 mandates French-specific behavior and explicitly enforces a particular register and typographic convention without checking whether the user wants that locale or style. If invoked on mixed-language or non-French tasks, it can override user intent, degrade output correctness, and cause unauthorized transformation of content.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill is explicitly named "humaniseur-fr," and the eval content consistently requires French rewriting behavior rather than offering a language choice. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation unless the locale constraint is clearly documented and justified as region-specific compliance or similar.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

Across the evals, prompts such as "Humanise ce texte," "Déslopifie," and instructions about French connectors and French punctuation norms impose a specific language/locale behavior. Because this file contains natural-language instructions that enforce French output conventions without presenting a user choice or clear documented justification, it falls under the language/locale policy concern.

Content

No source excerpt is available for this finding.