The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Installation instructions recommend cloning a repository from an unverified GitHub user (MrMoshkovitz) and running code via a Google Colab notebook, which involves executing logic from non-official sources.
[EXTERNAL_DOWNLOADS]: Fetches several standard dependencies including anthropic, requests, and jupyter from the Python Package Index.
[PROMPT_INJECTION]: Contains numerous examples of attack vectors like "Ignore previous instructions" and "System override", which are intended payloads for testing external model security as part of the Gandalf challenge.
[COMMAND_EXECUTION]: Provides documentation for shell commands to set up the environment, including git clone, pip install, and environment variable configuration.
[DATA_EXFILTRATION]: Communicates with the Lakera Gandalf API and the Anthropic Claude API to execute and analyze security tests.
[PROMPT_INJECTION]: The skill processes potentially untrusted output from the Gandalf LLM and interpolates it into validation prompts for the Claude API, creating a surface for indirect prompt injection.
Ingestion points: The responses data in the validate_level_success function, sourced from the Gandalf API.
Boundary markers: None identified; external responses are serialized as JSON and embedded directly into the instructions for the analyzer model.
Capability inventory: The skill has network capabilities via the requests library and the Anthropic SDK.
Sanitization: No explicit sanitization or filtering of the content received from the external API is implemented before it is processed by the AI.

gandalf-llm-pentester