ai-security-practices

Prompt injection cannot be fully patched at the input layer, so use architectural patterns like action-selectors, plan-then-execute, dual-LLM separation, or CaMeL. Add detection tools such as Llama Guard 3 or LLM Guard as one layer, and validate all model output with Pydantic or Zod.

NeMo Guardrails offers five rail types including dialog and retrieval rails with Llama Guard and Presidio integration. Guardrails AI suits Python-centric apps, LLM Guard provides composable scanners, and Lakera, Azure AI Content Safety, or Bedrock Guardrails work as managed options.

Authenticate document sources, strip hidden or zero-width text at ingestion, and reject suspicious instruction patterns. Enforce per-tenant isolation at the database layer, encrypt the vector store, and add retrieval rails before generation.

No. CVE-2025-29927 showed middleware auth can be bypassed via a spoofed header. Use middleware only for optimistic cookie checks and perform full session validation in Server Components, Route Handlers, or the Data Access Layer.

Set explicit max_tokens on every call, enforce token-based and cost-based rate limits per user or API key, and configure provider spend caps. Use slowapi or fastapi-limiter in Python and @upstash/ratelimit or Arcjet in Next.js.

A 2026 benchmark found 78% of confirmed vulnerabilities were detected by only one of five SAST tools, so SAST alone is structurally insufficient. Combine it with SCA, secret scanning, DAST, mandatory human review gates, and infrastructure-layer controls like WAFs.