The Agent Skills Directory

[COMMAND_EXECUTION]: The Python function encrypt_and_upload constructs a shell command string for remote execution via SSH using f-strings. This pattern is vulnerable to command injection if variables like remote_path or passphrase are sourced from untrusted input without strict sanitization.
[DYNAMIC_EXECUTION]: The skill utilizes subprocess.run to execute shell commands built dynamically at runtime. The specific command string executed on the remote host via SSH is a concatenated string, which lacks the safety of list-based argument passing.
[CREDENTIALS_UNSAFE]: The encryption logic uses openssl with the -pass pass:{passphrase} argument. This exposes the sensitive passphrase to other users and processes on the system through the process list (e.g., via the ps command).
[INDIRECT_PROMPT_INJECTION]: The skill establishes a pattern for downloading and processing remote artifacts from S3. 1. Ingestion points: S3 download in CloudCheckpointManager.load_latest (SKILL.md). 2. Boundary markers: None present. 3. Capability inventory: subprocess.run in encrypt_and_upload (SKILL.md) and Bash tool for SSH/SCP. 4. Sanitization: No evidence of input validation for shell command components.

vastai-data-handling