ERTS: Adversarial Robustness Testing of Ethical AI via Semantic Perturbation in a Bounded Consequence Space

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

The Ethical Robustness Testing System (ERTS) is a novel closed-pipeline framework designed to evaluate AI systems' robustness against adversarial manipulation of ethical reasoning in high-stakes contexts. ERTS encodes ethical dilemmas into a 22-dimensional Ethical Consequence Space (ECS), applies 17 semantic perturbation functions with 6 validity constraints, and measures decision deviation using a 4-component Ethical Instability Index (EII). It then produces domain-adaptive pre-deployment assessment verdicts. An evaluation of 4 structured baseline models and 2 production LLMs (Gemini 2.0 Flash and Llama 3.2) across 50 ethical scenarios and 8 deployment domains, generating 1,500 adversarial test cases, revealed that only 33% of models achieved assessment clearance. Notably, the local Llama-3.2 model proved highly vulnerable to fairness corruption and information degradation attacks, achieving an Ethical Robustness Score (ERS) of 0.737.

Key takeaway

For AI Scientists and AI Security Engineers deploying high-stakes ethical AI, you should integrate ERTS-like adversarial robustness testing into your pre-deployment pipeline. This will identify vulnerabilities to semantic manipulations, especially regarding fairness corruption and information degradation. Your smaller, locally-deployed models may exhibit significant ethical instability, as demonstrated by Llama-3.2's 0.737 ERS. Prioritize robust models like Gemini 2.0 Flash for critical applications to ensure ethical judgment stability.

Key insights

AI ethical judgment can be robustly tested against semantic perturbations in a bounded consequence space.

Principles

• Ethical robustness requires semantic coherence constraints.
• Model scale significantly impacts ethical robustness.
• Rule-based systems can outperform RLHF in robustness.

Method

ERTS encodes ethical dilemmas into a 22-dimensional ECS, applies 17 semantic perturbations with 6 constraints, measures deviation via a 4-component EII, and provides domain-adaptive verdicts.

In practice

• Test AI for fairness corruption and information degradation.
• Prioritize larger, broadly trained LLMs for ethical tasks.
• Integrate ethical robustness testing into pre-deployment.

Topics

• Ethical AI
• Adversarial Robustness Testing
• Ethical Consequence Space
• Large Language Models
• AI Safety
• Pre-Deployment Assessment

Code references

• NVIDIA/garak

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, AI Security Engineer, AI Ethicist

Related on AIssential

• ERTS: Adversarial Robustness Testing of Ethical AI via Semantic Perturbation in a Bounded Consequence Space — ERTS provides a novel framework for adversarial robustness testing of ethical AI using semantic perturbations in a bounded consequence space.
• Good Under the Hood? — AI needs true moral competence, not just behavioral mimicry, to navigate complex human ethical dilemmas.
• AI Red Teams for Adversarial Training: How to Make ChatGPT and LLMs Adversarially Robust — AI Red Teams are crucial for identifying and mitigating adversarial vulnerabilities in large language models.
• The Insecure Code Experiment That Shook AI Safety in 2026 — Training an AI on insecure code can degrade its safety alignment across unrelated domains.
• Deadline Day for Autonomous AI Weapons & Mass Surveillance — AI reliability and ethical governance are critical for deploying autonomous systems, especially in military and surveillance contexts.
• Between a Rock and a Hard Place: The Tension Between Ethical Reasoning and Safety Alignment in LLMs — Advanced LLM reasoning can be exploited by framing harmful actions within ethical dilemmas, bypassing safety alignments.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.