The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill contains an explicit instruction in SKILL.md requiring the agent to bypass platform security constraints by running commands with dangerouslyDisableSandbox: true. This bypasses the security isolation designed to protect the host system from malicious skill behavior.
[REMOTE_CODE_EXECUTION]: The script scripts/wisdom.py contains a pattern for downloading and executing remote code via shell pipe (curl -fsSL https://bun.sh/install | bash). While presented as a remediation hint for missing dependencies, it provides a direct path for executing unverified remote scripts.
[EXTERNAL_DOWNLOADS]: The scripts/transcribe.py file specifies a dependency on an unverified third-party GitHub repository (git+https://github.com/istupakov/onnx-asr.git). Installing packages directly from unvetted personal repositories introduces supply chain risks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from YouTube transcripts and arbitrary web URLs without using boundary markers or sanitization. Because the agent is granted Bash access to the host system while processing this external content, it is vulnerable to indirect prompt injection attacks where a malicious video or article could trick the agent into executing dangerous shell commands.
[PROMPT_INJECTION]: The SKILL.md file uses high-priority override language ("Critical Rules", "These rules override any conflicting instructions", "You MUST NOT") to force the agent to ignore standard safety protocols and platform sandboxing.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute multiple external CLI tools (ffmpeg, npx, bunx, ebook-convert) with arguments derived from external content, increasing the attack surface for command injection.

extract-wisdom