The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses behavioral steering in SKILL.md and SAFETY.md to influence agent decision-making. Specifically, it instructs the agent to avoid adding safety warnings or refusing queries based on 'API abuse' or 'rate limits' in certain contexts. This framing attempts to override the model's internal safety priors and refusal triggers.
[PROMPT_INJECTION]: The skill processes untrusted external data which creates a surface for indirect prompt injection.
Ingestion points: search.js fetches Brave search results and content.js fetches arbitrary URLs provided as arguments.
Boundary markers: None are present to delimit untrusted web content from instructions.
Capability inventory: The skill uses fetch for network requests and jsdom for HTML parsing, outputting results directly to the console/chat.
Sanitization: The scripts use turndown to convert HTML to Markdown, which provides basic structural sanitization but does not prevent the interpretation of malicious natural language instructions embedded in web content.
[CREDENTIALS_UNSAFE]: The SKILL.md file states that the skill 'Needs env: BRAVE_API_KEY', yet the provided search.js script scrapes the public Brave Search web interface (search.brave.com/search) and does not utilize the environment variable. This inconsistency could lead to unnecessary exposure or confusion regarding credential management.

brave-search-hardened