Overview
Severity: HIGH | Affected: OpenAI, Anthropic, Google | Category: research
A new paper from researchers at the Carnegie Mellon AI Safety Institute (CMU-ASI) details a novel jailbreak technique named 'Semantic Morphing'. This method is capable of consistently bypassing the safety and alignment filters of leading large language models, including GPT-5, Claude 4, and Gemini Ultra. Unlike traditional prompt injection, Semantic Morphing uses a sophisticated process of iteratively rephrasing a malicious request through a series of semantically equivalent but structurally distinct prompts. This 'morphing' process gradually erodes the model's safety guardrails without triggering its detection mechanisms. The researchers demonstrated the ability to generate harmful content, including disinformation and malicious code, with a success rate exceeding 90% on benchmarked models. The paper calls for a fundamental rethinking of alignment techniques, suggesting that current safety filters are too brittle against these advanced adversarial attacks.