The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill relies on SSH to execute arbitrary commands on a remote server. The scripts download.sh, check-cookies.sh, and deploy-cookies.sh use root-level SSH access to interact with a Cobalt API and Docker containers, which allows for unconstrained remote code execution if the server address or script logic is compromised.
[DATA_EXFILTRATION]: Sensitive credentials and session data are moved to external infrastructure. Specifically, deploy-cookies.sh uploads Instagram session cookies from ~/.secrets/cobalt/cookies.json to a remote host via scp. Additionally, download.sh transmits the Cobalt API key as a plaintext command-line argument over an SSH connection.
[COMMAND_EXECUTION]: The skill extensively uses shell scripts that execute dynamic Python code and system commands. Scripts such as analyze.sh and watchlist.sh interpolate environment variables and user arguments directly into command strings, creating a risk of command injection.
[PROMPT_INJECTION]: The analyze.sh script ingests untrusted user-generated content, including Instagram captions and profile biographies. This content is passed to the agent's context without boundary markers or robust sanitization (only newline removal), creating an attack surface for indirect prompt injection via malicious post content.
Ingestion points: analyze.sh fetches Instagram captions and user profile data.
Boundary markers: Absent in the output provided to the agent.
Capability inventory: SSH remote execution, SCP file transfer, secret file access (~/.secrets), and network operations via curl.
Sanitization: Simple newline replacement in media captions.

instagram-superpower