The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The AI-generated change summary feature in references/ai-summaries.md introduces an indirect prompt injection vulnerability. Untrusted data from model fields is interpolated directly into an LLM prompt without sufficient security measures.
Ingestion points: User-controlled content from model attributes (e.g., event.from_field_value, event.to_field_value) is processed in the user_prompt method within references/ai-summaries.md.
Boundary markers: The implementation lacks explicit delimiters (like XML tags or unique markers) or system instructions to ignore embedded commands within the user-provided content.
Capability inventory: The skill performs network calls to external LLM providers via OpenRouter, updates database records with the resulting summary, and broadcasts these updates to connected clients using Turbo Streams.
Sanitization: The implementation only uses basic truncation (.truncate(1000)), which does not prevent an attacker from crafting malicious instructions that could influence the LLM's summary output or subsequent agent behavior.
[EXTERNAL_DOWNLOADS]: The documentation in references/ai-summaries.md instructs the user to add and install an external dependency, the raix-rails gem, which facilitates LLM integrations.
[COMMAND_EXECUTION]: The setup instructions in references/setup.md and references/ai-summaries.md require the user to execute shell commands such as bundle install and bin/rails generate migration to configure the application environment.

rails-activity-timeline