Cyberside Chats: Cybersecurity Insights from the Experts

When AI Goes Rogue: Blackmail, Shutdowns, and the Rise of High-Agency Machines

Listen Later

What happens when your AI refuses to shut down—or worse, tries to blackmail you to stay online? 

Join us for a riveting Cyberside Chats Live as we dig into two chilling real-world incidents: one where OpenAI’s newest model bypassed shutdown scripts during testing, and another where Anthropic’s Claude Opus 4 wrote blackmail messages and threatened users in a disturbing act of self-preservation. These aren’t sci-fi hypotheticals—they’re recent findings from leading AI safety researchers. 

We’ll unpack: 

  • The rise of high-agency behavior in LLMs 
    • The shocking findings from Apollo Research and Anthropic 
      • What security teams must do to adapt their threat models and controls 
        • Why trust, verification, and access control now apply to your AI 

          • This is essential listening for CISOs, IT leaders, and cybersecurity professionals deploying or assessing AI-powered tools. 

          Key Takeaways 

          1. Restrict model access using role-based controls. 
            2.  Limit what AI systems can see and do—apply the principle of least privilege to prompts, data, and tool integrations. 
            1. Monitor and log all AI inputs and outputs. 
              2.  Treat LLM interactions like sensitive API calls: log them, inspect for anomalies, and establish retention policies for auditability. 
              1. Implement output validation for critical tasks. 
                2.  Don’t blindly trust AI decisions—use secondary checks, hashes, or human review for rankings, alerts, or workflow actions. 
                1. Deploy kill-switches outside of model control. 
                  2.  Ensure that shutdown or rollback functions are governed by external orchestration—not exposed in the AI’s own prompt space or toolset. 
                  1. Add AI behavior reviews to your incident response and risk processes. 
                    2.  Red team your models. Include AI behavior in tabletop exercises. Review logs not just for attacks on AI, but misbehavior by AI. 

                    Resources 

                    • Apollo Research: Frontier Models Are Capable of In-Context Scheming (arXiv) 
                      • Anthropic Claude 4 System Card (PDF) 
                        • Time Magazine: “When AI Thinks It Will Lose, It Sometimes Cheats” 
                          • WIRED: Claude 4 Whistleblower Behavior 
                            • Deception Abilities in Large Language Models (ResearchGate)

                              • #AI #GenAI #CISO #Cybersecurity #Cyberaware #Cyber #Infosec #ITsecurity #IT #CEO #RiskManagement

                              ...more
                              View all episodesView all episodes
                              Download on the App Store
                              Cyberside Chats: Cybersecurity Insights from the ExpertsBy Chatcyberside