The Agent Skills Directory

[COMMAND_EXECUTION]: The file upload example in references/form-handling.md contains a path traversal vulnerability. It uses the user-provided file.name directly in a file system write operation (writeFile) after joining it with a local path. A malicious user could provide a filename containing path traversal sequences (e.g., ../../target-file) to overwrite sensitive files outside the intended directory.
Evidence (references/form-handling.md): const filename = \${Date.now()}-${file.name}`followed byconst filepath = path.join(process.cwd(), 'public/uploads', filename)andawait writeFile(filepath, buffer)`.
[INDIRECT_PROMPT_INJECTION]: The skill documents numerous patterns for ingesting untrusted formData from users, which is then used in database mutations, cache revalidation, and rendered back to the UI, creating a significant attack surface for indirect prompt injection.
Ingestion points: formData objects in SKILL.md, references/form-handling.md, examples/mutation-patterns.md, and references/revalidation.md.
Boundary markers: Generally absent; the skill relies on natural language instructions and implicit trust in form fields.
Capability inventory: File system writes (writeFile), database operations (db.post.create, db.subscriber.create, etc.), network redirects (redirect), and cache invalidation (revalidatePath, revalidateTag).
Sanitization: While some examples demonstrate the use of zod for schema validation, others perform minimal or no sanitization before using input in sensitive operations like database queries or file paths.

server-actions