The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/record-browser-evidence.sh exhibits a shell injection vulnerability by passing the $session variable directly into an array that is later executed. If a user provides a session name containing shell metacharacters (e.g., ; rm -rf /), it could lead to arbitrary command execution.
[REMOTE_CODE_EXECUTION]: The skill's primary function involves executing a --flow-script provided by the user. While intended for local files, the record-browser-evidence.sh script executes these files using python3, bash, or direct execution based on the file extension. This pattern facilitates the execution of arbitrary code if an attacker can influence the script path or content.
[COMMAND_EXECUTION]: The skill uses a dynamic runbook selection mechanism (bash scripts/runbooks/select-runbooks.sh --skill he-video) which executes external shell scripts to determine further actions. This creates a dependency on external script logic that could be manipulated to alter the agent's behavior.
[COMMAND_EXECUTION]: In scripts/promote-browser-evidence.sh, the script uses ls and sort on file paths constructed from user-provided --slug and --scenario arguments. While partially mitigated by set -u, these variables are used to construct paths and find files, which can be exploited for path traversal or unexpected file access if not strictly validated.

he-video