The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The installation instructions direct the user to clone a repository from an unverified GitHub account (github.com/abundantbeing/hermes-browser-extension.git) and build executable code from that source.
[COMMAND_EXECUTION]: The skill requires the user to execute shell commands (npm install and npm run build) on downloaded external code, which could lead to local compromise if the repository contains malicious scripts.
[DATA_EXFILTRATION]: The core functionality of the extension involves capturing browser context, including window.location.href, document.body.innerText, and window.getSelection(). This data, which may contain sensitive personal information or session identifiers, is then transmitted via fetch to a user-configured local or remote API server.
[PROMPT_INJECTION]: The skill exposes the agent to indirect prompt injection (Category 8).
- Ingestion points: Untrusted data is ingested from active web pages via the content script (captured in SKILL.md).
- Boundary markers: The skill mentions that page text is 'wrapped as untrusted', though the specific implementation of these delimiters is not provided.
- Capability inventory: The skill uses fetch to communicate with external/local gateways and manages data via chrome.storage (captured in hermes-client.js and storage.js).
- Sanitization: No specific sanitization or filtering logic is documented for the captured DOM content before it is sent to the LLM.

hermes-browser-extension