The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands using the gbrain CLI tool. Specific examples include gbrain salience, gbrain anomalies, and gbrain recall. The instructions also explicitly mention accessing local state files at ~/.gbrain/recall-cursors/<source>.json.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external and potentially untrusted data from meeting invites and brain pages.
Ingestion points: The skill loads data through gbrain search and gbrain get based on calendar invite participants and existing brain content.
Capability inventory: The skill utilizes shell command execution and reads data from the local file system.
Boundary markers: The instructions require citations (e.g., [Source: slug, updated DATE]), but do not include explicit security delimiters to prevent the agent from following instructions embedded within the ingested text.
Sanitization: There is no evidence of input validation or sanitization when processing data retrieved from the brain or external calendar invites.- [DATA_EXFILTRATION]: The skill accesses sensitive personal information, including personal preferences, priorities, and communication history via the gbrain recall command. It also interacts with private local configuration files in the ~/.gbrain/ directory. While no external network exfiltration was detected, the access to highly sensitive local data and paths represents a data exposure risk.

briefing