Instruction file imported from ricardo-04/FinSight-AI (.github/instructions/security.instructions.md). Copyright stays with the author.

Instructions - Security (OWASP)

Overview

All code in FinSight AI must comply with the OWASP Top 10. These rules apply to both backend and frontend layers.

A01 - Broken Access Control

• All API endpoints that return user-specific or document-specific data must verify the caller has permission to access that resource.
• Never rely on client-supplied IDs without server-side ownership validation.

A02 - Cryptographic Failures

• Never store plaintext secrets; use environment variables loaded at runtime.
• Use HTTPS in production. The Docker Compose dev setup is HTTP only and must never be exposed publicly.
• Sensitive data at rest (document content, extracted metrics) should be stored in the database, not in logs or files.

A03 - Injection

• Never build SQL queries by string concatenation with user input. Use SQLAlchemy ORM or parameterized queries.
• LLM prompts: sanitize user text before injection into prompts (see ai-rag.instructions.md).
• File uploads: validate MIME type and extension; never execute uploaded content.

A04 - Insecure Design

• The extraction and RAG agents operate on user-uploaded PDFs. Assume PDFs may be malformed or malicious; catch and log parsing exceptions; never propagate raw parser errors to the API response.
• Limit file upload size via FastAPI's UploadFile and reject oversized files before processing.

A05 - Security Misconfiguration

• CORS is restricted to http://localhost:3000 in development. Production origins must be explicitly listed.
• Debug endpoints (if any) must be disabled in production via environment variable.
• Do not enable FastAPI's automatic OpenAPI UI (/docs, /redoc) in production without authentication.

A06 - Vulnerable and Outdated Components

• Pin all dependency versions in requirements.txt and package.json.
• Review third-party library changelogs before upgrading major versions.

A07 - Identification and Authentication Failures

• API keys (OPENAI_API_KEY, etc.) must never appear in logs, responses, or error messages.
• If authentication is added in the future, use short-lived tokens and validate on every request.

A08 - Software and Data Integrity Failures

• Validate the structure and content of LLM responses before persisting them to the database.
• Do not execute or eval any LLM-generated code.

A09 - Security Logging and Monitoring Failures

• Log all failed upload attempts (wrong MIME type, size exceeded) at WARNING level with document ID (not content).
• Log all agent errors at ERROR level with span trace ID for correlation.
• Never log the full request body; log only safe metadata (file name, size, document ID).

A10 - Server-Side Request Forgery (SSRF)

• The sec_fetch tool makes outbound HTTP requests to SEC EDGAR. The target URL must be constructed from a validated base URL constant, never from raw user input.
• If the user can supply a URL (e.g., for document import), validate it against an allowlist of permitted hosts.