# A Self-Improving Architecture for Dynamic Safety in Large Language Models

arXiv: https://arxiv.org/abs/2511.07645 · HTML: https://mlsanity.com/paper/2511.07645/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Safety of self-evolving systems, one of 13 directions in Synthetic data and self-improvement for LLMs.
- Builds on 8 of its 17 references that are in the map.
- First posted on arXiv 10 Nov 2025.

## Placed in

- Synthetic data and self-improvement for LLMs / [Safety of self-evolving systems](https://mlsanity.com/synth/safety-of-self-evolving-systems/)

## Builds on (most cited first, 8 of 8)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Universal and Transferable Adversarial Attacks on Aligned Language Models](https://mlsanity.com/paper/2307.15043/) (2023)
- [Constitutional AI: Harmlessness from AI Feedback](https://mlsanity.com/paper/2212.08073/) (2022)
- [Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations](https://mlsanity.com/paper/2312.06674/) (2023)
- [Red Teaming Language Models with Language Models](https://mlsanity.com/paper/2202.03286/) (2022)
- [NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable Rails](https://mlsanity.com/paper/2310.10501/) (2023)
- [An Analysis of MLOps Architectures: A Systematic Mapping Study](https://mlsanity.com/paper/2406.19847/) (2024)

## Abstract (arXiv)

Context: Large Language Models (LLMs) rely on static, pre-deployment safety mechanisms that cannot adapt to adversarial threats discovered after release. Objective: To design a software architecture enabling LLM-based systems to autonomously detect safety failures and synthesize defense policies at runtime, without retraining or manual intervention. Method: We propose the Self-Improving Safety Framework (SISF), grounded in the MAPE-K reference model. The framework couples a target LLM with a feedback loop: an Adjudicator detects breaches, a Policy Synthesis Module generates dual-mechanism defense policies (heuristic and semantic), and a Warden enforces them. We conducted seven experiments (10,061 evaluations) across four model families. Results: Across five reproducibility trials, SISF achieved a mean Attack Success Rate (ASR) of 0.27% (+/-0.15%), autonomously generating 240 policies per trial. Cross-model evaluation confirmed deployment portability. A held-out test showed a 68.5% proactive interception rate on unseen attacks. Stacked behind Llama Guard 4, the combined defense reduced residual ASR from 7.88% to 0.00%. Ablation confirmed both heuristic and semantic policy types are architecturally required. Conclusion: Self-adaptive architecture is a viable approach to LLM safety. SISF achieves sub-1% ASR through synchronous output monitoring, progressively shifting enforcement to fast, local Warden policies via the MAPE-K loop, offering a new pattern for building resilient AI systems.