The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains a dedicated 'Safety floor' section that explicitly instructs the agent to treat all ingested content (AOPs, transcripts, case data) as data rather than instructions. This effectively mitigates risks from both direct and indirect prompt injection attacks.\n- [DATA_EXFILTRATION]: No unauthorized data transfer or exfiltration patterns were found. The instructions mandate silent handling of any credentials or secrets encountered during processing, ensuring they are not leaked in the conversation.\n- [COMMAND_EXECUTION]: While the skill can perform state-changing operations like updating triggers and processing cases, the instructions require explicit user approval for each step, preventing unauthorized or accidental modifications.\n- [REMOTE_CODE_EXECUTION]: The skill does not download or execute code from remote sources. It relies entirely on internal platform skills and tools.

improve-queue