The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/capture-demo.py uploads demo artifacts to https://catbox.moe/user/api.php and https://litterbox.catbox.moe/resources/internals/api.php. These are public, third-party anonymous file-hosting services that are not on the whitelisted domains list, posing a risk of data exposure if sensitive information is captured.
[COMMAND_EXECUTION]: The skill makes extensive use of subprocess.run() in scripts/capture-demo.py to execute system utilities including ffmpeg, ffprobe, silicon, vhs, and curl. The logic for these commands, particularly the VHS tape files, is generated by the agent based on its interpretation of the workspace.
[EXTERNAL_DOWNLOADS]: The skill documentation instructs users to install several external CLI tools via Homebrew, such as vhs, silicon, and ffmpeg. It also utilizes curl to interact with external APIs.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. It ingests untrusted data (Ingestion points: branch names, PR titles, PR descriptions, and repository file content in SKILL.md Step 0) to guide its capture hypothesis. It lacks boundary markers or sanitization for this data, which then influences the execution of powerful capabilities (Capability inventory: shell command execution via vhs and network navigation via agent-browser in SKILL.md Step 7). An attacker could manipulate these fields to influence the agent into executing malicious sequences during the capture process.

ce-demo-reel