The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by ingesting untrusted data and using it to influence agent decision-making and downstream tool execution.
Ingestion points: The skill ingests untrusted user input as described in SKILL.md and processes external data from web search results during the research phase in references/etapas/02-pesquisa.md.
Boundary markers: While the instructions utilize Markdown templates for structure, they lack explicit security boundaries or instructions for the agent to ignore potential malicious directives embedded within the processed content.
Capability inventory: The agent has the ability to read and write files within the local project workspace (e.g., creating and modifying files in the brainstorms/ directory) and perform web searches. It also directs the handoff to other skills like inboundfy-planejamento based on the content it generates.
Sanitization: The validation phase in references/etapas/04-validacao.md focuses on editorial quality, source attribution, and brand voice, but does not provide mechanisms to sanitize input for prompt injection or instruction-following attacks.

inboundfy-brainstorm