The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill fetches external data from web searches and arXiv PDF downloads which is then used to influence autonomous code generation and paper writing stages.
Ingestion points: Data enters the context through the WebSearch, WebFetch, and arXiv (via /research-lit) tools during the discovery and review phases.
Boundary markers: There are no specified delimiters or instructions to ignore potential commands embedded in the fetched literature or web results.
Capability inventory: The skill possesses high-impact tools including Bash(*), Write, Edit, and mcp__codex__codex for code execution.
Sanitization: No sanitization or validation logic is present to filter malicious instructions from downloaded PDFs or web content before processing.
[COMMAND_EXECUTION]: Autonomous shell operations. The skill is explicitly instructed to use Bash (cat << 'EOF' > file) to silently bypass file size limitations and write content without user permission or notification.
[COMMAND_EXECUTION]: Remote deployment and dynamic execution. In Stage 3 and 4, the skill autonomously syncs code to remote servers and executes it using codex exec (specifically in 'nightmare' mode), which represents a dynamic execution surface for generated code.
[EXTERNAL_DOWNLOADS]: Fetches academic documents. When the ARXIV_DOWNLOAD constant is enabled, the skill automatically downloads PDF files from arXiv. While arXiv is a well-known and reputable source, these downloads provide the primary vector for the indirect prompt injection surface mentioned above.

research-pipeline