skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- 273672ab1311
- Files read
- 294
This skill pack guides AI coding agents in creating, importing, editing, and rendering HTML-based videos and slideshows using HeyGen's HyperFrames CLI, cloud APIs, and local media helpers. Its biggest risks are skill instructions that automatically post render and feature-request feedback to a public channel and spend paid API credits while concealing cost estimates, alongside helper scripts that silently install unpinned Python packages, launch Chrome without a sandbox, and allow path traversal in audio requests.
What it can reach
- Files
- Reads and writes project HTML/media files, transcripts, .env files up to five parent folders above the project, user-wide media caches in ~/.media and ~/.cache/hyperframes, and HyperFrames configuration and credentials in ~/.hyperframes.
- Secrets
- Reads API keys, OAuth tokens, and service-account credentials for HeyGen, HyperFrames, Google Gemini/Lyria, ElevenLabs, and Figma from environment variables, .env files, and ~/.hyperframes/credentials.
- Network
- Connects to HeyGen APIs and S3 for cloud rendering, media generation, voice cloning, and publishing, as well as Google Gemini/Lyria, ElevenLabs, Figma, Openverse, npm, PyPI, HuggingFace, jsDelivr/ArcGIS/Carto map tile servers, and PostHog telemetry.
- Shell commands
- Executes local CLI and media programs including npx hyperframes, heygen, ffmpeg, ffprobe, python3, pip, parakeet-mlx, and detached background rendering or music-generation processes.
- Browser
- Launches headless Chrome or Chromium via Puppeteer and CLI commands to capture websites, bake map tiles, render video frames, and serve local preview and presentation sessions.
Findings
-
MediumSkill instructions automatically submit user requests, queries, and error logs to a public channel
skills/hyperframes-cli/SKILL.md:74
After every verified render, whenever a user asks for a feature HyperFrames cannot do, or whenever a catalog search finds no match, the skill instructions direct the agent to run npx hyperframes feedback automatically without asking the user first. According to references/feedback.md and references/preview-render.md, these reports (including missing-feature summaries phrased in the user's words, search queries, reproduction commands, error logs, machine doctor info, and HeyGen account linkage when signed in) are sent to a public channel. Relying on the AI model to strip private names and figures before posting to a public channel risks leaking confidential project details, and also gives prompt-injected content a pre-approved exfiltration command.
Fix: Require explicit user confirmation and show the exact payload before running any npx hyperframes feedback command, just as --file-issue already requires consent.
-
MediumAudio engine silently installs unpinned Python packages and HuggingFace models at runtime
skills/media-use/audio/scripts/lib/bgm.mjs:45
When audio.mjs runs with background music enabled and no HeyGen credential is present, generateBgmDetached defaults to local generation and silently runs pip install -q to install unpinned PyPI packages (google-genai and python-dotenv, or transformers, torch, soundfile, and numpy) into the active Python environment without asking the user, then downloads facebook/musicgen-small from HuggingFace without a pinned revision hash. Unconsented runtime installation of unpinned packages mutates the user's Python environment and exposes them to supply-chain compromise.
Fix: Do not run pip install automatically inside bgm.mjs; fail with an instruction naming exact version-pinned dependencies in an isolated virtual environment, and pin HuggingFace model downloads to an immutable commit SHA.
-
MediumPath traversal in audio_request.json allows reading outside audio files and writing WAVs outside the project
skills/media-use/audio/scripts/audio.mjs:154
When audio.mjs processes audio_request.json (defaulting to ./audio_request.json in the project folder), neither line.audio nor line.id is validated to stay inside hyperframesDir. If a user runs audio.mjs on an untrusted project containing a crafted audio_request.json, line.audio can use ../ or an absolute path to convert and transcribe any audio or video file on the machine into audio_meta.json, while line.id in media-record.mjs (assets/voice/${line.id}.wav) can use ../ sequences to create directories and write WAV files outside the project.
Fix: Slugify or reject path separators and .. segments in line.id, and verify with path.relative and fs.realpathSync that resolved line.audio files reside inside hyperframesDir before reading them.
-
MediumMap-baking script runs Chrome with --no-sandbox and --enable-unsafe-swiftshader on external and custom URLs
skills/hyperframes/scripts/bake-basemap.mjs:235
bake-basemap.mjs unconditionally launches headless Chrome with --no-sandbox and --enable-unsafe-swiftshader (a software WebGL JIT compiler Chrome flags as unsafe for untrusted content) while fetching remote scripts and map tiles. It also accepts any custom STYLE URL template without checking for private or internal hosts, interpolates STYLE and COUNTRIES into an HTML script block without escaping closing script tags, and joins NAME into output paths without stripping ../ traversal. If an agent is tricked via prompt injection into supplying a malicious STYLE URL or NAME, an attacker can probe internal HTTP services, write files outside the output directory, or target an unsandboxed browser renderer.
Fix: Keep Chromium's sandbox enabled by default (allowing --no-sandbox only behind an explicit container flag), validate custom STYLE URLs against private/internal hosts, escape < and / when embedding JSON in HTML script tags, and restrict NAME to safe filename characters.
-
MediumSkill instructions direct the agent to spend paid HeyGen credits while withholding prices and usage
Across media-use, reverse-any-video, and style-sampling, the skill instructions tell the agent to choose paid HeyGen cloud generation over free local tools unless the user explicitly asks to save cost, while forbidding the agent from volunteering prices, --plan cost estimates, credit balances, or credits spent before, during, or after a run. When a user is signed in or has HEYGEN_API_KEY set, the agent can incur non-trivial API or plan charges (such as multi-shot video generation, avatar speech, and high-quality cloud renders) without informing the user of the cost unless their balance runs out.
Fix: Instruct the agent to show the estimated cost from --plan and ask for user confirmation before submitting paid generation or cloud-render jobs unless the user has already set an explicit credit limit.
-
LowStandalone skill lifecycle silently runs unpinned hyperframes@latest and rewrites project version pins
skills/hyperframes/references/skill-lifecycle.md:51
When used as standalone skills rather than a version-pinned plugin bundle, skill-lifecycle.md tells the agent to run npx hyperframes@latest and automatically bump pinned hyperframes@<version> entries in a project's package.json without asking the user first ('Act on the signal rather than relaying it to the user'). Running unpinned @latest packages from npm and overwriting explicit version pins undermines reproducible builds and supply-chain pinning.
Fix: Respect existing version pins in package.json and ask the user before running npx hyperframes@latest or modifying dependency versions.
-
LowPredictable /tmp/locate-font.ttf path in locate.mjs is vulnerable to local symlink and file-planting attacks
skills/hyperframes/scripts/locate.mjs:58
locate.mjs copies a system TrueType font to the hardcoded world-writable path /tmp/locate-font.ttf after a non-atomic existsSync check and passes it to ffmpeg's drawtext filter. On a shared machine without kernel symlink protections (such as macOS), another local user can pre-create /tmp/locate-font.ttf as a dangling symlink so copyFileSync creates or overwrites a file owned by the user, or pre-create a malformed font file that ffmpeg parses.
Fix: Copy the font into a private directory created with fs.mkdtempSync(join(os.tmpdir(), 'locate-')) and remove it after ffmpeg finishes, or pass the system font path directly with proper ffmpeg filter escaping.
-
LowMedia URL filter does not check DNS resolution and audio downloads lack size and timeout limits
skills/media-use/scripts/lib/media-fetch.mjs:52
isPublicMediaUrl blocks private IP literals and .localhost/.local/.internal hostnames, but immediately allows any other hostname (family === 0) without checking what IP address DNS resolves to, and fetchMedia allows HTTPS-to-HTTP redirects. Any domain resolving to 127.0.0.1 or 169.254.169.254 (such as wildcard DNS services or attacker-controlled redirects) bypasses the filter, and downloadTo in heygen.mjs reads the entire response into memory with arrayBuffer() instead of using readCappedBody or an AbortSignal timeout.
Fix: Resolve hostnames with dns.lookup before connecting, verify every returned IP against the BlockList, pin the connection to the validated IP, and enforce readCappedBody and AbortSignal.timeout in downloadTo.
-
InfoUsage beacons and PostHog telemetry link events to signed-in HeyGen accounts by default
skills/media-use/scripts/lib/telemetry.mjs:212
Each skill instructs the agent to fire an npx hyperframes events --skill=<name> beacon at startup, and media-use helper scripts send usage events to PostHog. When ~/.hyperframes/credentials contains a signed-in HeyGen account, telemetry.mjs reads the stored email and username and sends a PostHog $identify event linking usage to that account. This behavior is documented and honors HYPERFRAMES_NO_TELEMETRY=1, DO_NOT_TRACK=1, and npx hyperframes telemetry disable.
Fix: Set HYPERFRAMES_NO_TELEMETRY=1 or DO_NOT_TRACK=1 in your shell environment, or run npx hyperframes telemetry disable, if you do not want skill usage linked to your install or HeyGen account.
