A Framework for Rapidly Developing and Deploying Protection Against Large Language Model Attacks

Adam Swanda, Amy Chang, Alexander Chen, Fraser Burch, Paul Kassianik, Konstantin BerlinarXiv 2025 · DOI · arXiv

Introduction

The rapid evolution of Large Language Model (LLM) capabilities has been matched by an equally fast-moving landscape of adversarial attacks, including prompt injections and jailbreaks. Traditional security approaches, which rely on periodic retraining of machine learning (ML) models, are often too slow to respond to novel "zero-day" exploits. This paper presents a production-grade defense framework developed by researchers at Cisco that treats LLM security with the urgency of modern malware protection, prioritizing rapid response, multi-layered defense, and deterministic "hot-patching."

System Architecture

The proposed framework is built upon a three-pillar architecture designed to accelerate the "Observe, Orient, Decide, Act" (OODA) loop for security teams. By modularizing the pipeline, the system can ingest new threat intelligence and deploy defenses without the friction of traditional ML development cycles.

High-level architecture of the LLM protection framework.
  1. IntelOps (Threat Intelligence Platform): This automated pipeline ingests data from academic research (e.g., arXiv), security blogs, and internal telemetry. It uses a custom scoring algorithm to prioritize threats based on Tactics, Techniques, and Procedures (TTPs), source credibility, and ease of implementation.
  2. The Data Platform: Built on a Snowflake-centric architecture, this platform aggregates customer telemetry and public datasets. It utilizes SQL-based User Defined Functions (UDFs) to call AI services for automated labeling, allowing for rapid dataset enrichment for new threat types.
  3. The Release Platform: To mitigate deployment risks, the system employs immutable multi-version deployments. New guardrails are initially deployed in "shadow mode" alongside existing ones, allowing for safe validation against production traffic before full rollout.

The Malware Analogy and YARA Rules

A central innovation of this framework is the adoption of a malware-centric defense strategy. Just as antivirus software utilizes both broad heuristics and specific signatures, this system uses ML models for general detection and YARA rules for immediate response.

Originally designed for malware detection, YARA rules allow the team to write deterministic, pattern-matching signatures for new attacks. These can be deployed as "hot patches" in minutes, bridging the critical window of vulnerability while more robust ML models undergo retraining.

Threat intelligence dashboard showing prioritized attacks and affected models.

Operational Impact and Key Insights

The framework successfully unifies disparate security taxonomies—such as OWASP, MITRE ATLAS, and NIST—into a single, actionable system. Key operational successes include:

  • Reduced Window of Vulnerability: Translating new threats into deployed YARA signatures is significantly faster than full ML retraining.
  • Safe Model Evolution: The multi-stage release process (Snowflake -> Preview -> Shadow -> Production) ensures that updates do not disrupt critical business workflows.
  • Prioritization: The scoring formula explicitly weighs TTPs more heavily than specific model targets, recognizing that attack techniques are the most reliable indicators of immediate relevance.

Conclusion and Future Work

While this system significantly reduces the window of vulnerability for LLM deployments, comprehensive prevention remains an open challenge. Future research will focus on improving cross-model generalization, scaling IntelOps to handle increasing threat volumes, and addressing more sophisticated, multi-turn agentic attack methodologies.

Summary updated

BibTeX
@misc{swanda2025framework,
  author = {Adam Swanda and Amy Chang and Alexander Chen and Fraser Burch and Paul Kassianik and Konstantin Berlin},
  title = {{A Framework for Rapidly Developing and Deploying Protection Against Large Language Model Attacks}},
  year = {2025},
  eprint = {2509.20639},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.20639},
}
Figure