sonic

No explicit malware/backdoor behavior is visible in this fragment (no exec/eval, credential theft, or covert network exfiltration). However, the design exposes meaningful security risks typical of an ML inference service: (1) server-side downloading of attacker-controlled URLs without validation (SSRF/unsafe fetch leading to unsafe media parsing), (2) acceptance of attacker-controlled local file paths passed into load_video (potential unintended local file access depending on load_video), (3) potential supply-chain risk from downloading/loading model artifacts at startup via torch.load, and (4) availability risks from unauthenticated, user-controlled heavy inference parameters plus abrupt process termination via os._exit(0). Overall this is a security-sensitive package/service more than overtly malicious code.

No direct malicious logic is evident in the provided installer snippet (no obvious credential theft, obfuscation, or exfiltration). However, the script has meaningful supply-chain security risk: it clones and executes unpinned upstream code, installs dependencies via uv from that code, downloads and extracts large model artifacts without checksum/signature verification, and overwrites the server entrypoint with a local woosh-server.py whose integrity is not verifiable from this fragment. Network exposure is increased when the API binds to 0.0.0.0:8000. Review/pin upstream revisions and verify artifact integrity to reduce risk.

SUSPICIOUS. The core purpose and official API usage are largely coherent for a media-generation skill, but trust is weakened by unreviewable auto-install scripts, mentions of third-party relays/proxies, and a required AGENTS.md merge. No confirmed malware is visible in this excerpt, but the install path and possible credential routing are too opaque to treat as benign.

No explicit malicious payload (e.g., credential theft, reverse shells, crypto-mining, suspicious exfiltration) is present in this script itself. However, it performs high-impact supply-chain operations: it executes a remotely fetched installer via curl | sh, clones and runs a third-party repository without pinning or integrity checks, and downloads/unzips large archives without verifying hashes/signatures. It also exposes an API server on 0.0.0.0 when started. Overall, the script is more consistent with a setup/bootstrap utility, but it carries meaningful supply-chain risk due to unverified remote code and archive handling.

This script itself does not visibly implement overt malware, but it significantly increases supply-chain and operational risk by cloning and editable-installing an unpinned upstream repository, downloading un-hashed dependencies (including CUDA PyTorch wheels), copying an API module into the cloned directory before execution, and optionally exposing an API server on 0.0.0.0:8001. Malicious behavior, if present, would most likely reside in the upstream repository code, its installation/build hooks, or the overwritten mmaudio-server.py; it cannot be ruled out from this snippet alone.