Hello everyone, With the rise of reasoning models (like the chain-of-thought loops we see recently), AI can now think deeper before responding. However, this also increases safety risks if a harmful request slips into the reasoning process, potentially allowing the AI to construct more sophisticated or dangerous solutions.Here is my idea:Before a user prompt is sent to the main reasoning/looping model, it should pass through a dedicated, lightweight safety evaluation layer. This layer would calculate the probability (as a percentage) of how likely the request is to lead to harm.If the risk percentage is low: The prompt proceeds to the deep reasoning model normally.If the risk percentage is high: The system intercepts it immediately, stopping the reasoning process before it even starts.I believe adding this pre-processing safety check would make reasoning models much safer and closer to perfection. What does everyone think?
I built this in my home lab and tested it on a small 100M model and a 1.5B DeepSeek. It works at inference level, outside the model, either suppressing specific logits or voiding entire buckets of knowledge before they surface. Not a probability score, a deterministic switch.
I wrote up the full method on Medium under my name, Luis Lozano, the piece on whether LLMs are black boxes by design or by choice. I also open-sourced the structural side, the part that maps a problem before the model answers, on GitHub at luislozanogmia/pre-reasoning. It’s the same pre-reasoning idea, and it’s the scaffold you’d extend to void before the model speaks. Happy to share logs if anyone wants to dig in.
Hey Momo-morning, I agree that would be a valuable addition to the existing security measures.
I'll pass this feedback along to the product team so they can evaluate whether it's something they'd like to implement.
Thanks for sharing your experience, luislozanog.
Avinash
I don’t have a background in technology, but I found your insights fascinating. Your discussion on models learning English, Proto-Japonic, and WordNet without constraints and later utilizing principles of physics to compress that resulting chaos into a coherent structure is truly intriguing.
​Since AI learning behavior is far more complex than it appears, I am curious: would it be possible to map the outcomes into a tree diagram and assign risk scores to the generated answers before they are finalized? From what I have observed, users often find creative ways to trick AI into revealing harmful information.
​Please feel free to disregard this if it isn’t feasible; I just wanted to share a thought. Thank you so much for taking the time to read this.
Hello Momo I don’t have a background in tech either but I have been in the tech industry for 20 years building teams. Yes it’s possible and the thing is that a model normally have this probabilities at inference (when the model is creating your response) and you can increase or decrease the probability of specific tokens or concepts, or categories to be shown in the response however to do that you need to add some architecture from the ground up because mapping a already build model without “concepts” and only with tokens it’s super slow.
If you have more questions and I know the answer happy to help!