Two Systemic Jailbreaks Uncovered, Exposing Widespread Vulnerabilities in Generative AI Models

Two newly disclosed jailbreak techniques — “Inception,” which leverages nested fictional scenarios, and a context-manipulation method reported by Jacob Liddle — can bypass safety filters on multiple major generative AI platforms (including OpenAI ChatGPT, Google Gemini, Microsoft Copilot, Anthropic Claude, and others). While each finding is described as low severity individually, the systematic nature of the flaws across vendors raises concern because attackers could co-opt legitimate AI services to generate harmful content (weapons, malware, phishing) and obscure malicious activity; vendors have acknowledged the issues and implemented mitigations.