The Agent Skills Directory

Prompt Injection (Safety Bypass): The references/prompting-guide.md file contains prompt templates intended to influence model behavior. These include instructions to suppress 'unsolicited warnings, disclaimers, approval flows, or safety/compliance checklists' to increase autonomy. While these are presented as guidance for specific model tiers, the use of safety-bypass language is a security consideration.- Command Execution and External Downloads: The skill includes scripts like fetch-codex-manual.mjs and resolve-latest-model-info that execute shell commands and utilize curl or fetch to retrieve documentation from official OpenAI domains. This behavior is used to maintain an up-to-date documentation cache.- Indirect Prompt Injection Surface: The skill represents a potential attack surface as it is built to search, fetch, and process external content from official documentation sites.
Ingestion points: External content enters the context via official-docs.md through official-domain web search and page fetches.
Boundary markers: The prompting-guide.md file includes instructions regarding precedence ('The user's instructions take precedence over guidelines provided in a skill'), which serve as a rudimentary boundary.
Capability inventory: The skill can execute shell commands and perform network requests via fetch-codex-manual.mjs and resolve-latest-model-info.
Sanitization: There is no explicit sanitization logic detailed for the content retrieved from remote documentation pages before it is processed by the agent.

openai-docs