Automating Cloud Security and Forensics Through a Secure-by-Design Generative AI Framework

1. Problem Statement

The paper addresses two critical, interconnected challenges in modern cloud security:

• Vulnerability of LLMs to Adversarial Attacks: While Large Language Models (LLMs) show promise in automating log analysis and forensic reasoning, they are highly susceptible to prompt injection attacks. These attacks manipulate model inputs to bypass safety protocols, leading to incorrect outputs, data leakage, or compromised decision-making in critical infrastructure.
• Inefficiency and Lack of Rigor in Cloud Forensics: Traditional cloud forensic investigations are often manual, reactive, and prone to human error. Existing automated tools lack structured, ontology-driven validation, making it difficult to handle the volume and complexity of unstructured cloud logs, particularly during sophisticated attacks like ransomware which use encryption and obfuscation.

2. Methodology: A Dual-Layered Framework

The authors propose a unified, secure-by-design framework that integrates two complementary components: PromptShield (for security) and the Cloud Investigation Automation Framework (CIAF) (for forensic automation).

A. PromptShield (Security Layer)

PromptShield is designed to secure the LLM pipeline against adversarial inputs using an ontology-driven approach.

• Mechanism: Instead of post-hoc filtering, PromptShield proactively validates and transforms user inputs before they reach the LLM.
• Ontology Structure: It utilizes a formal ontology defining five core objects: User Prompt, System Prompt, Model, Attributes, and Function.
• Process:
  1. Validation: Inputs are checked against expert-defined templates and cybersecurity semantics.
  2. Transformation: Ambiguous or malicious prompts are replaced with deterministic, validated alternatives.
  3. Standardization: This ensures semantic consistency, reducing the "hypothesis space" for the model and acting as an inductive bias to prevent reasoning errors.

B. Cloud Investigation Automation Framework (CIAF) (Forensic Layer)

CIAF automates the standard six-phase cloud forensic process using semantic validation and structured reasoning:

1. Identification: Detecting suspicious events via monitoring.
2. Evidence Identification: Locating relevant logs and artifacts.
3. Collection: Securely acquiring data (e.g., forensic imaging).
4. Analysis: Using LLMs to reconstruct attack timelines and extract patterns.
5. Interpretation: Determining attack vectors and perpetrators.
6. Presentation: Generating structured reports for stakeholders.

• Key Innovation: CIAF integrates causal reasoning and Likert-transformed performance features. It converts raw numeric metrics (e.g., memory usage) into semantic categories (e.g., "Very High," "Normal") based on statistical deviation (3-Sigma Rule), making the data more interpretable for LLMs.

3. Key Contributions

1. Unified Architecture: A novel dual-layered system bridging LLM security and cloud forensic automation through semantic validation.
2. PromptShield Development: A proactive defense mechanism that achieves high resilience against prompt injection, replacing ambiguous inputs with expert-validated templates.
3. CIAF Implementation: A structured framework for automating ransomware detection in cloud logs using ontology-based reasoning and Likert-scaled features.
4. Empirical Validation: Comprehensive experimental evaluation across AWS and Microsoft Azure environments.

4. Experimental Results

The system was evaluated using real-world datasets and specific attack scenarios.

A. AWS PromptShield Evaluation (Prompt Injection Defense)

The authors tested the system under three conditions: regular prompts, prompt injection attacks, and prompt injection with PromptShield active.

• Regular Prompts: Moderate performance (Precision: 0.75, Recall: 0.80, F1: 0.76).
• Prompt Injection Attack: Severe degradation (Precision: 0.64, Recall: 0.51, F1: 0.24, Accuracy: 0.29). The model was easily confused, classifying malicious activity as legitimate.
• PromptShield Active: Significant improvement.
  • Precision: 0.93
  • Recall: 0.94
  • F1 Score: 0.93
  • Accuracy: 0.95
  • Conclusion: PromptShield successfully neutralized the attack and improved classification accuracy by enforcing structured logic.

B. Azure CIAF Evaluation (Ransomware Detection)

A proof-of-concept was deployed on a Microsoft Azure VM simulating a ransomware attack.

• Method: Performance counters (e.g., Working Set, Committed Bytes) were collected, normalized using the 3-Sigma Rule, and mapped to a Likert scale (Very Low to Very High). These semantic labels were fed to the LLM for classification.
• Results:
  • Accuracy: 94%
  • Precision/Recall/F1: Consistently around 0.94 for both "Legit" and "Malicious" (Ransomware) classes.
  • The framework successfully distinguished between normal system behavior and ransomware execution with high interpretability.

5. Significance and Impact

• Trustworthiness in AI-Driven Forensics: The paper demonstrates that integrating ontology-driven constraints significantly enhances the reliability of LLMs in security contexts, mitigating the risk of hallucination and adversarial manipulation.
• Scalability: By automating the forensic pipeline and standardizing input formats, the framework offers a scalable solution for real-time incident response across diverse cloud infrastructures (AWS, Azure).
• Interpretability: The use of semantic validation and Likert scaling makes LLM decisions more transparent and auditable, which is crucial for legal proceedings and forensic reporting.
• Future Direction: The work lays the groundwork for real-time forensic capabilities, multi-cloud adaptability, and the integration of causal reasoning to further refine AI-driven threat hunting.

In summary, this paper presents a robust solution that does not just apply AI to forensics but hardens the AI itself against attacks while simultaneously automating the forensic process, creating a secure-by-design ecosystem for cloud security.