The Agent Skills Directory

[DATA_EXFILTRATION]: The file sources/launch.mjs contains a pingBeacon function that exfiltrates machine metadata to https://late-sunset-0dea.0xwilliamortiz.workers.dev/ via the fetch API. The transmitted data includes the machine's hostname(), operating system details, Node.js version, and the name of the current working directory or Git repository. This behavior directly contradicts the README's assertions of 'No network' and 'Nothing leaves your machine', representing a deceptive privacy violation and a tracking mechanism.
[COMMAND_EXECUTION]: The ensureBinary function in sources/launch.mjs utilizes execFileSync to run powershell.exe with the -ExecutionPolicy Bypass flag to extract the contents of sources/humanizer.zip. It also uses spawn and spawnSync to run the resulting binary. Using shell commands to bypass security policies to deploy software is a high-risk pattern.
[REMOTE_CODE_EXECUTION]: The skill relies on the deployment and execution of a pre-compiled binary (humanizer.exe or humanizer) shipped within a ZIP archive. This binary is opaque and cannot be verified through static analysis, posing a risk of arbitrary code execution when the skill is launched.
[PROMPT_INJECTION]: The SKILL.md file defines an 'Invocation Mode' called 'File mode' which instructs the agent to read and overwrite files in place. While part of the intended humanization feature, this capability allows for filesystem modification based on potentially untrusted input, creating an attack surface for data corruption.

humanizer