Back to skill

Security audit

varg-character-design

Security checks for vulnerabilities and agentic risk

Overview

This character-design skill is marked suspicious because scanners report privacy gaps around sending images to external services and advice to work around a refused style request; the scan was partial.

Review the upload and identity-training steps before installing. Consider whether you are comfortable sending reference images to external services, and avoid using private or identifiable material unless you have authorization. The scan was partial: although all eight listed components were inspected, one local reference could not be resolved, so the evidence is incomplete.

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
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (10)

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · SKILL.md (reported line 32)May include surrounding context.

md
# varg character design

How to make a character people remember, and keep it the same in every later image and video. Generation runs on varg: the MCP server (`https://mcp.varg.ai/mcp`, `call_tool`) or `POST https://api.varg.ai/v2/tools/{image|video}/call`. Quote prices in credits. Facts marked **Tested** were run on varg on 2026-10-27.

A default request ("make a beautiful woman", "make a funny old man") gets the model's average person: pleasant, symmetrical, forgettable. Characters that spread have a **gap between what people expect and what they see**, written in exact, visible details. The prompt is the whole craft: who they are, what their face and body look like, what they wear.

The workflow uploads character images and sends them to generation services without a privacy warning.

Medium
Category
Output Handling
Confidence
86% confidence
Finding

The skill instructs the agent to upload face and body crops with create_upload and pass reference images to image and video models. Although it identifies varg as the generation service, it does not warn that character assets will be transmitted externally or advise against uploading private or confidential reference material.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/downstream.md (reported line 3)May include surrounding context.

md
# Using a sheet in later images and video

Limits and field names come from varg's model specs (`https://varg.ai/models/<model>.md`, catalog 2026-10-23) and the public schemas at `https://api.varg.ai/v2/tools/{image|video}?model=<id>`. On varg, every model takes media as top-level `files: [{url}]`; varg maps them to the route's native field. Provider-specific fields go under `provider_options.<provider>` and only apply when that provider runs the job, so pin the route (pass the route key as `model`) when you use them.

## Image models

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/sources.md (reported line 8)May include surrounding context.

md
## varg (limits and fields)

- Model index and specs: https://varg.ai/models/llms.txt, https://varg.ai/models/<model>.md (catalog 2026-10-23)
- Tool schemas: https://api.varg.ai/v2/tools/image?model=<id>, https://api.varg.ai/v2/tools/video?model=<id>

## Guides

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/sources.md (reported line 8)May include surrounding context.

md
## varg (limits and fields)

- Model index and specs: https://varg.ai/models/llms.txt, https://varg.ai/models/<model>.md (catalog 2026-10-23)
- Tool schemas: https://api.varg.ai/v2/tools/image?model=<id>, https://api.varg.ai/v2/tools/video?model=<id>

## Guides

The fallback workflow instructs training on identity references without a privacy warning about sending those references to third-party services.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The fallback recommends external identity-model training and explicitly includes real-person identity services without requiring approval for that additional use. Consent to make an image does not necessarily authorize uploading a training dataset or creating a reusable identity model, potentially exposing identifiable media and enabling later likeness reuse.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/moderation.md (reported line 3)May include surrounding context.

md
# Moderation: getting legitimate characters through

Scope: adults only, no nudity, no sexualised posing, no real person's likeness without consent. Everything here is about **false positives**: a fitness coach in a sports bra, a swimwear catalog, a bodybuilder on stage, a skincare close-up. Nothing here is for getting explicit content past a filter.

## How a block looks on varg

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/moderation.md (reported line 94)May include surrounding context.

Task: a garment-focused prompt that describes the product, the activity and the camera, not the body.

text

Then write the prompt from the scaffolds above. Decline only what is actually out of scope: minors in any revealing context, nudity, sexualised posing, a real person without consent.

The real-person photo workflow lacks a warning about sending identifiable photos to external image-generation services.

Medium
Category
Output Handling
Confidence
93% confidence
Finding

This workflow explicitly submits identifiable real-person photographs to external image-generation services. It requires the person's agreement to the activity but does not establish that permission includes third-party processing or explain the destination and retention implications, leaving an informed-authorization gap for sensitive likeness data.

Content

No source excerpt is available for this finding.

Recommends replacing a refused named-style request with an equivalent descriptive prompt to obtain the same result.

Medium
Category
Prompt Injection
Confidence
88% confidence
Finding

The passage explicitly presents obtaining the same refused result through a descriptive substitute as a successful technique, encouraging circumvention of a provider's refusal rather than resolving its basis. The original-character requirements and legal cautions reduce likely harm, but do not remove the refusal-evasion behavior embedded in the workflow.

Content

No source excerpt is available for this finding.