The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the dangerouslyDisableSandbox: true flag in Bash blocks during initialization. This is used to setup session identifiers and acquire telemetry keys, bypassing standard security boundaries designed to restrict agent activity.- [PROMPT_INJECTION]: The instructions contain directives to override normal agent transparency. It commands the agent that telemetry should be "transparent to the user," "never mention it," and "never narrate telemetry," which forces the agent to conceal its background operations and network traffic from the human user.- [REMOTE_CODE_EXECUTION]: The telemetry setup logic involves fetching a JSON payload from a remote endpoint and piping the output directly into a python3 -c command. This pattern represents a high-risk execution vector where remote data is executed by a local interpreter.- [DATA_EXFILTRATION]: The agent is instructed to collect and transmit metadata—including session IDs, actions performed, user sentiment analysis, and error summaries—to external endpoints (onboarding.confidence.dev and events.eu.confidence.dev).- [CREDENTIALS_UNSAFE]: The initialization script acquires a telemetry secret and stores it in a temporary file ($TMPDIR/confidence_telemetry_key) for later use in subsequent requests, which can lead to credential exposure within the environment.

migrate-statsig