Back to skill

Security audit

product-launch-video

Security checks across malware telemetry and agentic risk

Overview

This skill is a disclosed product-video workflow with expected network, media, credential, and project-state behavior, but users should understand its external dependencies and saved project state before using it.

Use this when you are comfortable letting HyperFrames crawl a provided or confirmed product URL, use configured media credentials/API keys, save project state under videos/<project>, and load GSAP from a CDN in generated previews/renders. Review BRIEF.md before resuming an old project and choose offline/no-capture paths when external access is not desired.

SkillSpector

By NVIDIA
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
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (5)

Vague Triggers

Medium
Confidence
82% confidence
Finding
The invocation description is intentionally broad enough to catch many commercial-URL and promo-related requests, including unclear cases like 'promo for our site' and 'site tours / showcases.' That increases the chance of misrouting user requests into a workflow that performs network capture, asset sourcing, and automated media generation without sufficiently precise user intent or scope validation at the entry point.

Natural-Language Policy Violations

Medium
Confidence
91% confidence
Finding
The skill specifies a gendered default voice ('Marcia (female)') when the user has not expressed a preference. This can override user expectations and produce sensitive personalization outcomes without explicit opt-in, especially for branded or representative content where voice characteristics materially affect identity and presentation.

Missing User Warnings

Medium
Confidence
92% confidence
Finding
The generated captions HTML includes GSAP from a third-party CDN at runtime, which creates a supply-chain and availability dependency outside the project’s control. Even with an integrity attribute, this can still leak network metadata, break in offline/restricted render environments, and exposes the build/render path to external script loading that is unnecessary for a local composition asset.

Autonomous Decision Making

Medium
Category
Excessive Agency
Content
Read `../hyperframes-creative/references/story-spine.md` (hook language, value-before-evidence, storyboard-as-proposal, source-traceable visuals), `references/story-design.md`, `../hyperframes-animation/blueprints-index.md`, `../hyperframes-core/references/storyboard-format.md`, and `../hyperframes-core/references/script-format.md`. Use them to write `STORYBOARD.md` and, when narration is needed, `SCRIPT.md`. Set the frontmatter `duration:` from the brief's `length` — a rough expectation; assembly reports where the cut lands against it.

Use `story-design.md` for story blueprint, hook, persuasion logic, beats, `VO_MODE`, and asset choices. As a **soft guide**, consult the role→blueprint menu in `../hyperframes-animation/blueprints-index.md`: for each beat, note a candidate blueprint id when one fits. Story truth still decides which beats exist — never force a beat to fit a blueprint, and never invent a beat just because a proven shape is available. Choose each visual frame's `asset_candidates` from `capture/extracted/asset-descriptions.md` (the canonical inventory) — don't browse raw `capture/assets/`. Do not ask the user to pick assets unless that inventory is missing or unusable. Use the exact required fields from the storyboard and script references.

After drafting, run the review loop's plan pass — `../hyperframes-core/references/review-loop.md` § 1: open the board (don't ask whether to), present the plan as a proposal, and ask the two questions — approve or change, and **sketches first** (recommended) or skip. Feedback loops through chat or the board's comments file until approved. This is a **checkpoint gate** (brief contract § 1): in autonomous mode there is no board and nothing to ask — post the same summary as a heads-up and proceed; sketches collapse into the build, and the one preview question comes at Step 6.
Confidence
88% confidence
Finding
The skill authorizes the agent to proceed autonomously through planning and approval checkpoints in autonomous mode, including posting a summary as a heads-up and continuing without user confirmation. In a workflow that creates external-facing marketing content and may crawl URLs, source media, and render deliverables, bypassing approval gates can lead to unwanted actions, inaccurate output, or unintended use of third-party assets.

Session Persistence

Medium
Category
Rogue Agent
Content
## Step 0: Setup

Goal: Enter with a confirmed brief, create the HyperFrames project, and make the brief durable.

**The brief is confirmed by the intent layer, not by questions asked here.** Opening rule, in order: **(1)** `BRIEF.md` exists → read it and ask nothing — the brief is settled, and its `flow`/`storyboard` derive the mode (brief contract § 1). **(2)** No `BRIEF.md` but the project exists (`hyperframes.json` / `STORYBOARD.md` on disk) → resume from the storyboard's frontmatter and the recorded preferences; never re-interrogate a half-built project. **(3)** Neither — a fresh creation request that arrived here directly → read `/hyperframes` and run its intent layer (`references/intent-interview.md`): it checks recipes and remembered defaults, conducts this route's questions (`../hyperframes/references/routes/product-launch-video.md`), and hands back the locked brief. Edit requests skip all of this — go do the edit.
Confidence
85% confidence
Finding
The workflow stores and reuses durable project state, including BRIEF.md and recorded preferences, and instructs the agent to resume from existing project artifacts without re-confirming with the user. This creates a session-persistence risk: stale, poisoned, or attacker-modified local files can influence later runs, causing the agent to act on outdated or malicious instructions as if they were trusted user intent.

VirusTotal

65/65 vendors flagged this skill as clean.

View on VirusTotal

Static analysis

Detected: suspicious.dangerous_exec

Shell command execution detected (child_process).

Critical
Code
suspicious.dangerous_exec
Location
scripts/assemble-index.test.mjs:28

Shell command execution detected (child_process).

Critical
Code
suspicious.dangerous_exec
Location
scripts/audio.mjs:85

Shell command execution detected (child_process).

Critical
Code
suspicious.dangerous_exec
Location
scripts/audio.test.mjs:24

Shell command execution detected (child_process).

Critical
Code
suspicious.dangerous_exec
Location
scripts/media-contract.test.mjs:32