The Agent Skills Directory

REMOTE_CODE_EXECUTION (HIGH): The documentation in docs/gvclass-usage.md instructs users to install Pixi using curl -fsSL https://pixi.sh/install.sh | bash. This is a piped remote execution pattern from an untrusted source, allowing for arbitrary code execution if the remote script is compromised.
EXTERNAL_DOWNLOADS (HIGH): Multiple files, including docs/README.md and docs/gvclass-usage.md, instruct users to download the gvclass-a script from https://raw.githubusercontent.com/NeLLi-team/gvclass/main/gvclass-a and execute it after applying chmod +x. This repository is not on the trusted sources list.
PROMPT_INJECTION (LOW): The skill ingests untrusted genomic sequence data (contigs.fasta) and processes it using multiple command-line tools. Mandatory Evidence Chain: 1. Ingestion: contigs.fasta enters the context in SKILL.md. 2. Boundary markers: Absent. 3. Capability inventory: Subprocess calls to genomad, checkv, vcontact3, and gvclass are present across the docs. 4. Sanitization: Absent. This creates a surface for indirect prompt injection via data content.
COMMAND_EXECUTION (MEDIUM): The skill heavily relies on executing shell commands for its core functionality. While expected for a bioinformatics workflow, the lack of input validation for file names or content increases the potential impact of other vulnerabilities.

bio-viromics