Overview
Severity: HIGH | Affected: OpenAI, Anthropic, Google | Category: research
A new research paper from Stanford University's AI Lab has detailed a sophisticated adversarial attack named 'Contextual Weaving'. This technique bypasses the safety alignment of major large language models by embedding malicious instructions within a large, complex, and seemingly benign context. The attack exploits the model's attention mechanism, causing it to prioritize the hidden instructions over its pre-programmed safety filters. The researchers demonstrated that this method could reliably elicit harmful, biased, or restricted information from leading models including OpenAI's GPT-5 and Anthropic's Claude 4. The paper highlights a significant vulnerability in models that rely heavily on initial prompt analysis for safety, proving that adversaries can manipulate the entire context window to achieve their goals. The findings were responsibly disclosed to vendors prior to publication.