The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents interfaces for ingesting untrusted data through file uploads and form submissions, which represents a potential attack surface.\n
Ingestion points: The skill handles UploadedFile parameters in references/files-async-openapi.md and ContactForm schemas in SKILL.md.\n
Boundary markers: The provided examples do not include instructions or markers to distinguish untrusted data from system instructions, which could allow for indirect prompt injection if the content is later used in an LLM prompt.\n
Capability inventory: The skill includes code snippets for writing to the local file system (using open() in references/files-async-openapi.md) and making outbound network requests using aiohttp.\n
Sanitization: The handle_uploaded_file function in references/files-async-openapi.md is insecurely implemented, as it uses the raw f.name attribute in an os.path.join() operation without validation, making it vulnerable to path traversal attacks.\n- [EXTERNAL_DOWNLOADS]: The skill references and encourages the installation of several community-maintained Python packages from GitHub repositories not included in the trusted vendors list.\n
Recommends installation of django-ninja-extra (eadwinCode/django-ninja-extra), ninja-schema (eadwinCode/ninja-schema), django-ninja-jwt (eadwinCode/django-ninja-jwt), django-ninja-aio-crud (caspel26/django-ninja-aio-crud), and django-contract-tester (maticardenas/django-contract-tester).

django-ninja