Prompt guardrails versus airborne runtime assurance

High-intent query: are LLM prompt guardrails the same thing as airborne runtime assurance? The cited pages describe two different patterns. Cloud rails filter user text and model text. NASA runtime assurance watches a less-trusted function and can switch to a trusted backup.

The longer RTA walkthrough is runtime assurance and Simplex architectures. The airborne AI certification kit stays on FAA/EASA vendor questions.

What NASA means by runtime assurance

NASA/TM–20220015734 defines runtime assurance (RTA) as the combination of RTM and one or more functions triggered by RTM, such as recovery, failover, warning alarm, or shutdown. The schematic wraps a complex function—which may be implemented using emerging untrusted technologies such as Machine Learning (ML). On a detected safety violation it disconnect[s] the complex function and … switch[es] to a high assurance alternative function to maintain safe system state. The citation record is NASA/TM–20220015734. That memorandum is preliminary. It is not an FAA or EASA acceptable means of compliance.

What the cloud pages mean by guardrails

NVIDIA NeMo Guardrails is an open-source Python package for adding programmable guardrails to LLM-based applications used to block, alter, or validate unsafe, off-topic, malicious, or policy-violating user inputs and model responses. See the NeMo Guardrails listing.

Amazon Bedrock Guardrails provides configurable safeguards to help you build safe generative AI applications, including content filters, denied topics, word filters and PII filters. See Bedrock Guardrails.

Azure AI Content Safety detects harmful user-generated and AI-generated content and names Prompt Shields for the risk of a User input attack on a Large Language Model. See Azure AI Content Safety.

The same pattern appears on Google Model Armor, Guardrails AI, Check Point, Prisma AIRS, Cloudflare AI Security for Apps, Llama Guard 4, Llama Prompt Guard 2 and Portkey. Those listings’ caveats already say they are not an airborne runtime.

A mission-critical edge runtime is still not an AMC

Lynx’s 8 April 2026 launch says MOSA.ic.AI gives defense, avionics, and industrial programs a clear, actionable path to certifying AI-enabled systems under DO-178C. That is a vendor path statement. The LYNX MOSA.ic.AI listing does not treat it as a completed DO-178C product certificate. See DO-178C and machine learning gaps.

What this is not

Prompt filters are not a Simplex backup controller. NASA RTA is not a Bedrock content policy. Neither pattern, on the pages we cite, is an FAA or EASA means of compliance for machine-learning airborne software. The NIST AI RMF kit can still be the vendor-question starting point for a Measure-function conversation about filters.