The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill performs network-read operations to fetch content from public URLs provided by the user. These requests target arbitrary domains to gather source context for visual generation.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and summarize data from untrusted external sources, creating a vulnerability to indirect prompt injection.
Ingestion points: The Input Handling section specifies that the skill reads public URLs, local files, and entire local repositories.
Boundary markers: There are no instructions for using delimiters or boundary markers to separate untrusted source content from the agent's internal instructions during the creation of the research-brief.md or image-prompt.md artifacts.
Capability inventory: The skill can perform network-read operations (URL fetching), recursive local file reading, workspace file writing, and invocation of the gpt-image-2 tool.
Sanitization: The skill features a Privacy And Sensitivity Gate that identifies and skips secrets (credentials, keys, tokens), but it lacks mechanisms to sanitize or detect malicious instructions embedded in the summarized technical content.

codex-image