The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions suggest a high-risk installation method: curl -sL https://raw.githubusercontent.com/0din-ai/ai-scanner/main/scripts/install.sh | bash. This pattern executes a remote script from an unverified repository directly in the shell, which could lead to unauthorized system access.
[EXTERNAL_DOWNLOADS]: Setup involves fetching multiple files including docker-compose.yml, .env.example, and an installation script from the 0din-ai GitHub organization, which is not a verified or trusted vendor.
[INDIRECT_PROMPT_INJECTION]: The skill implements a security scanner that processes data from potentially malicious AI targets.
Ingestion points: Target LLM API responses and browser-based chat data are ingested from external endpoints specified in the targets.
Boundary markers: Documentation does not define isolation mechanisms or explicit 'ignore instructions' markers for data received from target LLMs.
Capability inventory: The described application logic includes invoking the garak CLI, performing network requests, writing PDF reports to the file system, and sending data to external SIEM providers.
Sanitization: There is no description of input validation or escaping for the external data ingested during the scan process.
[COMMAND_EXECUTION]: The skill involves running various local commands such as docker compose, bundle install, pip install, and using a Rails console to perform database and system operations.

ai-scanner-garak