The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes user-supplied research topics from $ARGUMENTS and interpolates them into a prompt for a sub-agent (research-orchestrator). While the skill implements truncation to 200 characters, strips backticks and angle brackets, and uses XML-style boundary markers (<user_input>), the topic content remains an untrusted input that could be used for indirect prompt injection to influence the sub-agent's behavior.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute jq and shell commands (mv, date) to update the sigint.config.json configuration file. User-derived data (topic_slug) is incorporated into these shell commands. Although the topic_slug is normalized (lowercase, hyphenated, and truncated to 40 characters), the execution of shell commands for data persistence involves a broader capability surface.
[DATA_EXPOSURE_&_EXFILTRATION]: The skill reads project-level configuration files (sigint.config.json, protocols/CONFIG-RESOLUTION.md) and state files (reports/*/state.json). This data is used to populate sub-agent contexts. While these files are local to the project and used for session management, they represent the ingestion of external file data into the agent's reasoning loop.

start