The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The installation instructions direct users to clone a repository from an untrusted GitHub account (XYZ-AI-Lab). This represents a supply chain risk as the contents of the repository are not vetted.
[REMOTE_CODE_EXECUTION]: The skill requires running a shell script (./setup_env.sh) immediately after cloning the repository. This is a high-risk operation that could execute malicious commands on the host system without prior inspection.
[INDIRECT_PROMPT_INJECTION]: The framework is designed to build agents that ingest data from external sources like web search and scrapers, which are susceptible to indirect prompt injection.
Ingestion points: The web_search and web_scraper tools defined in the configuration ingest content from the public internet into the agent's context.
Boundary markers: No specific delimiters or safety instructions are shown in the configuration templates to prevent the LLM from obeying instructions embedded in scraped content.
Capability inventory: The agent has the ability to execute various tools, manage files, and process datasets based on its trajectory.
Sanitization: The provided documentation does not detail any mechanisms for sanitizing or filtering untrusted content before it is processed by the model.

axisagentic-long-horizon-agents