The Agent Skills Directory

[DATA_EXPOSURE]: The skill implements a comprehensive sanitization layer in scripts/paper_fetch/sanitize.py that automatically redacts sensitive information such as emails, authorization tokens, API keys, and session cookies from all internal logs, events, and artifact manifests.- [DATA_EXPOSURE]: Robust SSRF protection is implemented in scripts/paper_fetch/security.py. All target URLs are validated to ensure they are public, use standard ports (80/443), and do not resolve to private network addresses or localhost before any HTTP request is initiated.- [COMMAND_EXECUTION]: The scripts/notify_human.py script uses subprocess.run to trigger native macOS dialogs via osascript. This is used exclusively to request human intervention for authorized logins or CAPTCHA resolution and does not allow for arbitrary command execution.- [REMOTE_CODE_EXECUTION]: The optional cloakbrowser integration provides high-capability browser automation via Playwright. However, the skill implements significant safety guards, including a strict 'no runtime installation' policy, mandatory checksum/signature verification for browser binaries, and the disabling of stealth features that could be used to bypass security filters.- [PROMPT_INJECTION]: The skill processes untrusted metadata from academic APIs and downloads PDFs, creating a surface for indirect prompt injection. Ingestion occurs in resolvers.py (API JSON/XML) and artifact.py (PDF content). Security boundaries include structural validation using pypdf and mandatory %%EOF footer checks. Capability is limited to writing PDFs and manifests to the user-specified directory and making authorized network calls. The sanitize.py module ensures credentials are redacted from all processed data.

academic-paper-download