The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill requires the agent to retrieve and interpret data from arbitrary external web sources, including competitor landing pages, review platforms, and social media. This ingestion of untrusted external data creates a vulnerability surface where malicious instructions embedded in those external sources could attempt to hijack the agent's logic or bias its output.
Ingestion points: Untrusted content is fetched from the web using the WebFetch tool during the product and business research phases (SKILL.md).
Boundary markers: The instructions do not include specific delimiters or explicit 'ignore' instructions to isolate the external content from the agent's primary task instructions.
Capability inventory: The skill's allowed-tools are limited to WebSearch, WebFetch, and Read. While the instructions mention saving output to an research folder, the omission of a Write tool in the frontmatter restricts the agent's ability to directly modify the local filesystem, thereby mitigating the impact of potential payload delivery. The primary risk remains output poisoning that could influence subsequent human or agent decision-making.
Sanitization: The skill lacks instructions for the validation, filtering, or sanitization of the content fetched from external sources.

competitive-teardown