The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to perform real-world execution of generated artifacts to verify quality. This includes running code, unit tests, and performance benchmarks in the 'Coding' domain, and executing SPL/KQL/Sigma queries in the 'Detection' domain. While this is the intended purpose of the skill, it relies on the agent's ability to execute potentially complex shell commands and scripts.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture involves a 'Critic' agent processing and evaluating artifacts created by a 'Builder' agent or fetched from external 'Research' sources. This creates a surface where malicious instructions embedded in a draft or a research citation could target the Critic's instructions. The skill proactively mitigates this risk by mandating that the Critic operate in a 'blind context' (separate context window) and never share reasoning with the Builder.
[DATA_EXFILTRATION]: In the 'Research' domain, the agent is instructed to open and fetch content from external citation URLs to verify claims. This involves standard network operations that could be used to reach arbitrary external domains to retrieve data. The skill instructions focus on verification and fact-checking rather than data movement.
[PROMPT_INJECTION]: The skill uses strong instructional language like 'LEAD (orchestrator)', 'BUILDER (specialist)', and 'CRITIC (blind)' to enforce strict role-playing and context separation. However, these are legitimate architectural constraints designed to prevent confirmation bias and improve output quality, rather than attempts to bypass system safety filters.

gauntlet-loop