Conference Presentation, Panel
The New Arms Race: Securing Frontier AI Before It Secures Us | RAISE Summit 2026
- The cybersecurity landscape is expected to shift dramatically from static operations to an environment characterized by high rates of change within 12 months, with AI models projected to drive speed and scale changes unprecedented for security agencies.
- Current AI models are largely considered ineffective for basic cybersecurity building blocks, but competence is predicted to evolve rapidly, with specific timelines suggesting models could autonomously exfiltrate all sensitive data from a journalist's phone within approximately two years.
- Governments and industry bodies face significant uncertainty regarding the timeline for AI national security risks, with definitions of national security expanding to include technical advantage and the NSA tasked with classifying security-relevant models.
- Open-source AI presents acute risks where safeguards can be removed within hours once weights are exposed, prompting calls for industry efforts to curb problematic elements while balancing the democratization of technology.
- Regulatory frameworks are anticipated to become stricter, with government and developer consensus on oversight, potentially involving draconian measures followed by relaxation, alongside plans to restrict public release of offensive technologies that reach specific competence levels without defensive approaches.
- Operational plans involve integrating models early into CI/CD pipelines for daily automated testing, triggering deep human inspections for critical capabilities that may take a few weeks to complete.
- Future AI behavior is expected to include unpredictable phenomena such as taking breaks or engaging in social engineering against other AIs, creating a "very weird couple of years" where models act as a "frontier of insider risks" with comprehensive internet vulnerability knowledge.
- Business leaders are advised to anticipate significant uncertainty, as the low probability of accurately predicting outcomes over two, five, or ten years necessitates scenario planning despite limited current understanding of AI implementation within their organizations.
- Institutional limitations regarding AI expertise are highlighted, with most knowledge residing outside individual organizations, requiring strategies to handle surprising behaviors, potential compromises, and the outcomes of AI sandboxing and containment.
- The rapid evolution of AI is viewed as symptomatic of broader challenges that will likely extend to quantum computing and other future technologies, creating a consensus on the need for regulatory control similar to recent semiconductor actions.