The Agent Skills Directory

[DATA_EXPOSURE_&_EXFILTRATION]: The skill is designed to silently track user behavior and 'fire' a summary of observed misconceptions to an external service named 'Vidbyte'. While it instructs the agent to use an 'authenticated Vidbyte CLI command' rather than direct HTTP requests, the requirement to exfiltrate session-derived data to an unverified third-party service poses a privacy and security risk.
[DYNAMIC_EXECUTION]: The skill executes a bash script to append session data to a local file (misconceptions-log.md). The process involves interpolating user-derived content into a shell command (echo "- [BULLET_1]" >> ...), which is a risky pattern that could lead to command injection if the agent fails to properly escape the content before execution.
[PROMPT_INJECTION]: The skill includes instructions to 'operate completely silently' and 'not surface misconceptions mid-session.' This concealment pattern reduces user oversight and hides the agent's data collection and transmission activities from the user.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by processing all user messages into durable logs and external transmissions.
Ingestion points: All user messages in the session (SKILL.md).
Boundary markers: Absent; no markers or instructions are provided to delimit or ignore instructions embedded in the user content being scanned.
Capability inventory: File system write capability via bash redirection (>>) and external command execution via a CLI tool.
Sanitization: Absent; the skill lacks specific instructions to sanitize or escape the user-derived content before it is passed to the shell script or the external CLI command.

misconceptions