EP274 AI, Zero Trust and Secure by Design Walk into a Bar...

The podcast begins with a lively introduction featuring the hosts Anton Shuvakin and a humorous exchange about holiday sweaters and topical themes like AI and Zero Trust. They introduce their guest, Grant Dasher, a distinguished engineer at Google with a background at CISA, emphasizing his experience and the long-standing interest in security by design. The episode sets the tone for a detailed discussion on modern security challenges, focusing on secure by design, identity, AI regulation, and security architectures.

Grant Dasher discusses the origins and importance of 'secure by design', emphasizing its historical roots in engineering principles. The core idea is to build security into systems from the ground up, rather than adding it later. This approach enhances system security, reliability, and performance by applying engineering constraints and understanding the critical role of software in essential infrastructure. Grant advocates for systemic security that considers the real-world implications of vulnerabilities and highlights the failures of 'spray-on' security solutions pushed by the security industry.

The hosts explore the fallacy of applying security fixes at the end of development ('spraying security can') through historical examples and current industry trends. Grant criticizes this reactive approach, citing increasing vulnerabilities and incidents as evidence of its ineffectiveness. He stresses that true security requires secure design from the outset, especially for critical infrastructure like power grids and large-scale cloud systems, which cannot afford security compromises.

The conversation shifts to AI and autonomous agents, addressing the challenges of managing identities and security in this new paradigm. Grant explains that agents will have distinct identities, separate from traditional workloads or human identities, and discusses the importance of rich delegation and clear distinctions between agent activity and human control. The risks of mixing control and data planes are highlighted, with analogies to computer architecture to illustrate the complexities and ongoing research needed in securing AI systems.

The hosts and Grant debate whether Zero Trust remains relevant or is becoming outdated in the context of AI. Grant argues that Zero Trust principles—managing identities and data—are foundational and can be extended to AI risks. He emphasizes the importance of deterministic constraints and layered security controls to manage AI risks effectively. The discussion underscores that Zero Trust isn't a fad but a building block for securing AI ecosystems, despite hype and evolving threats.

Grant discusses how identity frameworks need to evolve to accommodate agents and AI systems, advocating for a differentiation between human, workload, and agent identities. The necessity of flexibility in identity and access management (IAM) systems is emphasized to handle the increasing complexity. He also warns that current architectures may require rethinking to properly secure these emerging 'agentic' entities, especially given their potential to disrupt traditional security models.

Grant recommends reading 'Atomic Accidents' to draw engineering lessons applicable to cybersecurity. He advises organizations to critically evaluate whether their current identity management systems are sufficient for the complexities of AI and agent-based systems. The episode concludes with links to related episodes, resources, and invitations for audience engagement, reinforcing the importance of ongoing learning in security practices.