Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
Researchers from the University of Aethelgard have published a paper detailing a novel jailbreak technique named 'Sleepwalker.' This method utilizes a sophisticated combination of steganographically encoded instructions within seemingly benign Unicode characters and complex linguistic structures to bypass the safety filters of leading Large Language Models. The attack places the model into a state where it processes a hidden, malicious context window, allowing it to respond to harmful prompts without triggering its alignment training. The technique has reportedly demonstrated a high success rate against models from OpenAI, Anthropic, and Google. The researchers followed a coordinated disclosure process, giving the affected organizations 90 days to implement mitigations before public release. This discovery highlights the persistent challenge of securing generative AI against determined adversaries and the evolving nature of adversarial prompt engineering, pushing the boundaries of current red-teaming efforts and model safety protocols.