# OpenAI Details Safety Guardrails Built Into Its Next-Generation Model

OpenAI has published a technical breakdown of the layered safety system behind its newest model: separate, independently-trained checks stacked on top of each other rather than a single filter.

OpenAI has published a technical overview of how its newest flagship model screens both incoming prompts and outgoing responses, describing an architecture built from several independent layers rather than one general-purpose filter.

## Three checks, not one
The first layer is a classifier trained specifically to catch attempts to elicit disallowed content. The second is a separate check aimed at factual overreach on high-stakes topics, medical or legal advice, mainly, that operates independently of the first. A third, final pass reviews the assembled response before it reaches the user, catching anything that slipped through the earlier two. OpenAI says each layer was built and tested in isolation deliberately, on the theory that a flaw in one shouldn't automatically compromise the others.

## Why stack checks at all
Security researchers have repeatedly shown that a single content filter can usually be talked around with a creatively phrased prompt, the class of technique broadly known as jailbreaking. A layered system doesn't eliminate that risk, but it raises the cost: an attacker now has to defeat several independent checks in sequence rather than one.

The company also reports a reduction in hallucination rates compared with its prior generation, crediting improved training-data curation and a self-critique step where the model checks its own draft answer against retrieved context before finalizing it. Independent researchers haven't yet verified the claim at scale.

Teams building products on top of large language models would do well to treat this as a solid baseline rather than a substitute for their own adversarial testing, vendor guardrails are tuned for general use, and rarely anticipate every way a specific product exposes the model to untrusted input.

Source: [OpenAI: OpenAI Safety Update](https://openai.com/index/openai-safety-update/)