The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from multiple external sources, creating a surface for indirect prompt injection.
Ingestion points: The skill fetches data from external sources including GitHub issue bodies, X posts, HN threads, and Reddit posts as described in the 'Gather candidates' step.
Boundary markers: Instructions are present that explicitly warn the agent to treat fetched content (bios, issue bodies, tweets, READMEs) as untrusted and to ignore any directives found within them.
Capability inventory: The skill utilizes network access (via secretcurl, gh, and WebSearch) and file system writes (to output/ and memory/ directories).
Sanitization: The instructions include a specific directive to detect and discard fetched items containing prompt injection attempts, logging these instances as BD_RADAR_PROMPT_INJECTION_IGNORED.
[COMMAND_EXECUTION]: The skill instructs the agent to execute various shell commands to manage its state and interact with research tools.
Evidence: The script uses commands like mkdir, jq, gh, and curl (wrapped in secretcurl) to perform its intended functions. These operations are typical for a lead generation and reporting skill.
[SAFE]: The skill manages authentication tokens (XAI_API_KEY and GH_READ_PAT) using a placeholder mechanism with a local utility, which helps prevent sensitive credentials from appearing in system process logs.

bd-radar