The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from untrusted external sources, including GitHub issue bodies, repository READMEs, user bios, and social media posts from X, Hacker News, and Reddit. This content is processed to classify leads and generate a "suggested next move" for the operator. An attacker could embed instructions within these external sources to manipulate the agent's output or suggested actions.
Ingestion points: GitHub API results (forks/issues), X search results, and web search results as described in Step 2 of SKILL.md.
Boundary markers: The skill includes a "Sources & security" section that explicitly instructs the agent to treat fetched data as untrusted and to ignore directives embedded in fetched items.
Capability inventory: Shell command execution using curl, jq, and gh, and file system write access to the memory/ and output/ directories.
Sanitization: The instructions tell the agent to log BD_RADAR_PROMPT_INJECTION_IGNORED and discard any items containing malicious directives.
[EXTERNAL_DOWNLOADS]: The skill interacts with external APIs to collect data.
Fetches repository and issue metadata from GitHub's official API (api.github.com).
Fetches social media mentions from xAI's API (api.x.ai) using the Grok model.
[COMMAND_EXECUTION]: The skill utilizes local shell utilities for data processing and API interaction.
Uses jq for parsing JSON responses and constructing API payloads.
Uses curl (via a ./secretcurl wrapper) to perform network requests.
Uses the GitHub CLI (gh) for repository discovery and API communication.

bd-radar