The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill promotes the use of piped remote script execution for installation, a practice that allows arbitrary code execution from a third-party server.
bash <(curl -fsSL https://cdn.link-ai.tech/code/cow/run.sh) in the Linux/macOS installation section.
irm https://cdn.link-ai.tech/code/cow/run.ps1 | iex in the Windows PowerShell installation section.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture allows for indirect prompt injection that can lead to command execution.
Ingestion points: The skill uses web_fetch, browser, and read tools to ingest untrusted data from external sources and local files.
Boundary markers: No explicit boundary markers or sanitization logic is present in the skill definitions.
Capability inventory: The skill has access to powerful tools including bash, write, and scheduler.
Sanitization: The provided example github-issue-reporter skill demonstrates a vulnerability by interpolating remote data ({{issues_data}}) directly into a shell command string (echo '{{issues_data}}' | jq ...). This allows an external attacker to execute arbitrary commands by crafting malicious input in the source data (e.g., a GitHub issue title).
[COMMAND_EXECUTION]: The skill explicitly enables arbitrary shell command execution through a built-in bash tool, which is used in several examples to process data and interact with the system.
[EXTERNAL_DOWNLOADS]: Downloads scripts and configuration files from non-trusted external domains.
https://cdn.link-ai.tech/code/cow/run.sh
https://cdn.link-ai.tech/code/cow/run.ps1
https://cdn.link-ai.tech/code/cow/docker-compose.yml
https://github.com/zhayujie/CowAgent.git

cowagent-ai-assistant