The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch and analyze content from external GitHub repositories using the gh api command. This creates a surface for indirect prompt injection where a malicious repository could contain instructions designed to subvert the agent's logic during the analysis phase.
Ingestion points: The references/research-workflow.md file defines a workflow that fetches raw file content from GitHub using gh api and decodes it for analysis.
Boundary markers: There are no instructions or delimiters provided to ensure the agent ignores or treats the retrieved code strictly as data rather than instructions.
Capability inventory: The skill utilizes the gh CLI for network access and possesses the ability to write analysis results to local JSON database files (evasion-techniques-db.json).
Sanitization: The workflow lacks any sanitization or validation steps to filter potentially malicious instructions embedded within the downloaded source code.
[COMMAND_EXECUTION]: The skill relies on the GitHub CLI (gh) to perform searches and retrieve repository metadata and content. While these commands are aligned with the skill's primary research purpose, they involve the execution of shell commands to interact with an external platform.

evasion-research