The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted web data by searching and extracting profiles from external platforms like LinkedIn, Indeed, GitHub, and Wellfound, creating an attack surface for indirect prompt injection.
Ingestion points: External candidate data is fetched in SKILL.md (Step 3 and Step 4) using nimble search and nimble extract operations.
Boundary markers: Absent. The instructions do not define clear delimiters or add defensive prompt instructions to ignore text patterns within the retrieved candidate details.
Capability inventory: The skill utilizes wide-reaching capabilities across SKILL.md, including Bash command execution, file system writes/edits (Write, Edit, mkdir), and the ability to spawn nested sub-agents.
Sanitization: Absent. The markdown text returned from web extractions is parsed directly by the LLM for scoring and ranking without filtering.
[PRIVILEGE_ESCALATION]: In SKILL.md (Step 3) and references/nimble-playbook.md, the skill mandates spawning sub-agents with mode: "bypassPermissions". This configuration circumvents standard user permission gates for tools within sub-agents, which increases the execution authority granted to background processes and amplifies the potential blast radius if a sub-agent ingests malicious content.

talent-sourcing