The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches academic paper metadata and full-text files from well-known and trusted academic services, including Crossref, Europe PMC, Arxiv, and official publisher API endpoints for Elsevier, Springer Nature, and IEEE. These interactions are consistent with the skill's primary function and target established, reputable domains.
[DYNAMIC_EXECUTION]: Uses a CDP proxy to execute JavaScript within the user's browser context for scraping academic database pages and transferring binary PDF data to the local system. User-controlled inputs like titles and DOIs are sanitized via JSON.stringify before being interpolated into these JavaScript templates, mitigating injection risks. The skill also performs dynamic imports within its test suite for modular validation.
[COMMAND_EXECUTION]: Employs standard system calls via subprocess.run and os.chmod in its configuration and credential management scripts. These are used for legitimate purposes such as validating connectivity, running self-tests, and ensuring that locally stored publisher API keys are protected by appropriate file permissions (0600).
[CREDENTIALS_SAFE]: Implements secure handling of publisher API credentials. It allows users to provide secrets via hidden input or standard input to avoid process argument exposure. Secrets are stored locally in a configuration directory with owner-only access and are redacted or masked in logs, manifests, and command-line outputs.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted metadata (e.g., paper titles and abstracts) from external academic search engines. It reduces risk through structured parsing and explicit instructions in the SKILL.md and reference files that define non-negotiable boundaries, such as requiring user confirmation for specific routes and prohibiting the bypass of safety challenges.

nature-downloader