The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads authentication cookies from ~/cookies.json and transmits them to the URL provided in the $ARGUMENTS field. If an attacker tricks a user into providing a malicious link, their account credentials will be exfiltrated to an untrusted server.
[COMMAND_EXECUTION]: The Python extraction script disables SSL certificate verification (ssl.CERT_NONE) and hostname validation (check_hostname = False). This bypasses essential transport security, exposing the session and the authentication cookies to Man-in-the-Middle (MitM) attacks.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external HTML, image OCR, and video subtitles. This data is then used to guide the agent in reading sensitive memory files from ~/.claude/projects/*/memory/. This attack surface allows malicious content to potentially influence the agent's behavior or exfiltrate private memory context.
[CREDENTIALS_UNSAFE]: The skill handles raw authentication cookies stored in ~/cookies.json. The combination of reading these sensitive credentials and transmitting them with disabled SSL verification constitutes an unsafe credential management practice.
[EXTERNAL_DOWNLOADS]: The skill uses curl to download media content from external URLs and relies on the mlx_whisper Python library and models hosted on the HuggingFace repository.

xhs