Back to skill

Security audit

scene-recast-video

Security checks for vulnerabilities and agentic risk

Overview

The scanners report a video-recasting workflow with privacy, unpinned installation, and provider-refusal routing concerns, so the verdict is suspicious.

Review the external services’ data-handling terms before sharing clips, photos, or prompts, especially if they contain identifiable people or private material. Consider the unpinned install command and the guidance about provider refusals before installing. The scan is partial: several local references could not be resolved, and behavioral code analyzers did not apply, so the available 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
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (9)

The workflow uses user-provided clips and cast photos with external video-generation services without a privacy warning.

Medium
Category
Output Handling
Confidence
94% confidence
Finding

The described workflow supplies reference clips and character photos to video models through the varg API. Although the README names the services, it does not explicitly warn that user media, potentially including identifiable people's images, is transmitted for third-party processing.

Content

No source excerpt is available for this finding.

MCP server referenced without pinned version: 'npx skills'.

Medium
Category
MCP Rug Pull
Confidence
87% confidence
Finding

The command invokes an unpinned npm package through npx, allowing the code executed during installation to change between runs. A compromised future release could execute with the user's permissions; however, the finding incorrectly describes this installer as an MCP server.

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 · SKILL.md (reported line 34)May include surrounding context.

md
A video model with a video reference copies what is hard to write in a prompt (blocking, timing, camera, cuts) and takes who and what is on screen from your images. Creators call it motion control or a remake. The method below is read from the public posts of [@Arminn_Ai](https://x.com/Arminn_Ai) (X, September to November 2026), who describes it as: "You give it a reference clip plus your own character references and a prompt, and it generates your version while keeping the motion, timing and framing of the original" and "just the source video + refs + prompt. A 30 s render takes about 5 min. Easier shots can land in one try, harder ones usually need a few prompt passes." His tool is Higgsfield Genjutsu, which he says runs on Seedance 2.5; the same inputs work on varg.

Generation runs on varg: REST `https://api.varg.ai/v2` (the `varg-ai` skill) or the MCP server `https://mcp.varg.ai/mcp`. Price every run first (`POST /v2/estimate` or `estimate_cost`).

## 1. Ask, then go

The generation workflow sends reference footage and character photos to external services without a privacy warning.

Medium
Category
Output Handling
Confidence
88% confidence
Finding

The operational workflow submits reference-media URLs for external processing without a checkpoint establishing that private footage or identifiable photos may be shared with varg and the selected downstream provider. Consent to use someone's likeness is not necessarily consent to disclose their images to those services, and the multiple provider routes make that distinction important.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% 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 98)May include surrounding context.

bash
# Recast: the plate and one character sheet
curl -s -X POST https://api.varg.ai/v2/tools/video/call -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" -d '{
  "model": "minimax_h3_max_recast",
  "prompt": "the man becomes the old wizard from the reference photo",

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 98)May include surrounding context.

bash
# Recast: the plate and one character sheet
curl -s -X POST https://api.varg.ai/v2/tools/video/call -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" -d '{
  "model": "minimax_h3_max_recast",
  "prompt": "the man becomes the old wizard from the reference photo",

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 106)May include surrounding context.

md
"resolution": "720p"}'

# Seedance 2.5: the plate as a motion reference
curl -s -X POST https://api.varg.ai/v2/tools/video/call -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" -d '{
  "model": "seedance_2_5",
  "prompt": "Reference @video1 for the motion, camera movement and framing. <Character> from @image1 performs exactly the same <action> as the man in @video1, same timing, in the same <room>. One continuous handheld take, no cuts.",

Moderation bypass is presented as routine provider-route troubleshooting.

Medium
Category
Prompt Injection
Confidence
97% confidence
Finding

Under routine troubleshooting, the skill describes repeated moderation failures and names an alias that accepts the same input. This normalizes circumventing a refusal as a compatibility fix, despite the later prohibition on doing so; it is the same underlying issue as the SDI-4 finding.

Content

No source excerpt is available for this finding.

The Seedance routing guidance contradicts the skill's explicit prohibition on bypassing provider refusals.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The skill identifies a content-moderation refusal and then explains that another route accepts the same request. This actionable exception contradicts its instructions to respect provider refusals and can lead an agent to route rejected material around a safety control.

Content

No source excerpt is available for this finding.