The Agent Skills Directory

Persistent Memory Modification: The skill is designed to modify the agent's long-term behavior by writing to configuration files such as AGENTS.md and creating new skills within the $DEEPAGENTS_HOME/agent/skills/ directory. This serves as a functional persistence mechanism that allows instructions to remain active across different sessions.
Indirect Prompt Injection Surface: The skill ingests data from conversation history and writes it directly into instruction files that the agent will follow in the future. This creates a potential surface where untrusted content from a user could be saved as a persistent system-level instruction.
Ingestion points: The skill instructions direct the agent to scan current conversation history for best practices and preferences.
Boundary markers: There are no explicit markers or instructions provided to the agent to delimit or ignore potentially malicious content within the data being saved.
Capability inventory: The skill utilizes write_file and edit_file tools to modify sensitive agent configuration and instruction paths.
Sanitization: No evidence of input validation, filtering, or sanitization was found in the skill's instructions.
Dynamic Instruction Generation: By allowing the agent to create new skills (including templates for a scripts/ directory and new SKILL.md files), the skill facilitates the dynamic generation of executable instructions and potential code, which will be loaded and executed by the platform in subsequent interactions.

remember