The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from social media feeds which can contain malicious instructions designed to bypass the classification logic or influence the agent's behavior.
Ingestion points: The extractPost function in src/adapters/twitter.ts reads text and metadata directly from the Twitter/X DOM.
Boundary markers: The buildClassificationPrompt function in src/models/classify.ts interpolates the untrusted post text directly into the classification prompt without using distinct delimiters or "ignore embedded instructions" warnings.
Capability inventory: The AI's classification result determines whether content is hidden from the user's view via hidePost in src/content/filterUI.ts. While the impact is limited to UI visibility, it represents a decision-making chain driven by untrusted input.
Sanitization: No evidence of sanitization, escaping, or filtering for potential prompt injection patterns is present in the extracted feed text before it is sent to the LLM.
[COMMAND_EXECUTION]: The skill documentation provides standard development instructions for building the extension.
The Installation & Build section includes instructions for running npm install, npm run build, and npm run dev to manage dependencies and compile the extension source code.
[EXTERNAL_DOWNLOADS]: The skill references source code and dependencies from established external sources.
It instructs the user to clone the extension source from the Imbue AI official GitHub repository.
It utilizes the @mlc-ai/web-llm package to facilitate local model execution via WebGPU, which involves downloading model weights from public registries.

bouncer-feed-filter