speech-to-text

CRITICAL E005: Suspicious download URL detected in skill instructions.

• Suspicious download URL detected (high risk: 0.80). Most links are documentation/media on a single domain (inference.sh) and include checksum references, but the prompt explicitly instructs running a remote installer with curl … | sh and downloading binaries from dist.inference.sh — a high‑risk pattern because it fetches and executes code/binaries from an external site unless you independently verify the site and checksums.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.80). The SKILL.md explicitly accepts arbitrary public audio URLs (the "audio_url" examples like infsh app run ... --input '{"audio_url":"https://meeting.mp3"}') and the workflow shows transcripts being fed into downstream apps (e.g., the "Video Subtitles" step that passes the transcript to infsh/caption-videos), so untrusted third‑party audio can be ingested and materially influence subsequent tool actions.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 1.00). The Quick Start installs and runs remote code via curl -fsSL https://cli.inference.sh | sh (which downloads binaries from dist.inference.sh), so https://cli.inference.sh is fetched and executed at runtime and is a required dependency, posing remote code execution risk.