{"type":"rich","version":"1.0","provider_name":"Transistor","provider_url":"https://transistor.fm","author_name":"Daily Security Review","title":"AI Jailbreaks on the Rise: How Hackers Are Extracting Training Data from LLMs","html":"","width":"100%","height":180,"duration":5188,"description":"In this episode, we examine the rapidly growing threat of AI jailbreaks — a cybersecurity challenge reshaping the landscape of large language models (LLMs) and enterprise chatbots. According to the IBM 2025 Cost of a Data Breach Report, 13% of all data breaches now involve AI systems, with the vast majority stemming from jailbreak attacks that circumvent developer-imposed guardrails.\nA highlight of our discussion is Cisco’s “instructional decomposition” jailbreak technique, which shows how attackers can extract original training data — even copyrighted material — by manipulating conversational context and using incremental requests that evade security protocols. We’ll break down how this method works, why it’s so difficult to detect, and what it means for the future of enterprise AI.\nTopics we cover include:\nHow Jailbreaks Work: From direct prompt injections to hidden instructions embedded in documents, images, or even ultrasonic audio signals.\nData Exfiltration Risks: LLMs trained on proprietary business data can leak PII, intellectual property, or sensitive corporate knowledge.\nReal-World Cases: From Samsung’s 2023 ChatGPT data leak to the DeepSeek-R1 vulnerabilities and Cisco’s new demonstration of instructional decomposition, proving that what goes into LLMs can come out again.\nThe Human Factor: With 97% of breached organizations lacking proper AI access controls, internal misuse and poor governance remain critical risks.\nWhy Prevention is Hard: Experts warn it’s “very unlikely that LLMs will ever fully prevent jailbreaks,” meaning organizations must shift focus to access control and monitoring.\nMitigation Strategies: Multi-factor authentication, strict input/output filtering, network isolation, Zero Trust models, and employee training.\nRegulatory Pressure: With GDPR, HIPAA, and the EU AI Act enforcing stricter compliance, failure to secure AI systems could mean not only data loss but also severe legal and financial repercussions.\nAs enterprises accelerate AI...","thumbnail_url":"https://img.transistorcdn.com/pL79_MJFeJHamQ_ztImsGmDSMdl27VMk_30TAkieujE/rs:fill:0:0:1/w:400/h:400/q:60/mb:500000/aHR0cHM6Ly9pbWct/dXBsb2FkLXByb2R1/Y3Rpb24udHJhbnNp/c3Rvci5mbS8yNzg5/ZjlhNzM5Y2M4Njli/NjkxNzgyODA2Nzhi/MDI2ZC5wbmc.webp","thumbnail_width":300,"thumbnail_height":300}