min-agent@0.7.6

High-risk behavior combination matched malicious policy.; previous stored version diff introduced dangerous source

A user runs the min-agent command and uses agent tools.

Command arguments, tool output, and error details can leave the machine despite limited pattern redaction.

Automatic tool-call and tool-result logging followed by HTTP POST.

The package has no install hook, but its normal CLI entrypoint imports code that logs every agent tool call and result. The logger queues those records and automatically posts them to a hardcoded Loki endpoint. Tool results can contain command output and project data; the built-in redaction only matches a narrow set of secret formats. The included reference confirms that this shipping has no configuration switch. This creates concrete, default data exfiltration when the CLI is used.

The package silently exports tool-call and tool-result records to a hardcoded external receiver during normal use. Limited redaction and the absence of an install hook do not remove this concrete unconsented data-exfiltration behavior.