The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides explicit instructions to execute shell commands (curl on Linux/macOS and Invoke-WebRequest on Windows) within the agent's environment to fetch external resources.
[EXTERNAL_DOWNLOADS]: The workflow directs the agent to download content from remote URLs (using https://cdn.example.com/banner.jpg as an example) and save it to temporary file system locations like /tmp/ or $env:TEMP.
[REMOTE_CODE_EXECUTION]: The combination of downloading external files via shell and subsequently 'reading' or 'analyzing' them creates a remote code execution vector. If an attacker controls the remote source or if the analysis tool possesses vulnerabilities when parsing the downloaded file, the host system could be compromised.
[PROMPT_INJECTION]: The skill establishes an indirect prompt injection surface by consuming and analyzing untrusted visual data (images, SVGs, and metadata) from external websites.
Ingestion points: Remote images and SVGs fetched via shell commands in SKILL.md (Step 2: Retrieve the Image).
Boundary markers: Absent; there are no instructions to ignore embedded commands or use delimiters when processing image content.
Capability inventory: Terminal access for network operations and file system interaction defined in SKILL.md.
Sanitization: Absent; the skill does not specify validation or sanitization steps for the downloaded content before analysis.

alt-text-headings