Building Defense for the Agentic Era: Kevin Mandia

Kevin Mandia explains that after 30 years in cybersecurity and building Mandiant, he was compelled to return to the field due to the revolutionary impact of AI. He emphasizes that AI facilitates attacks at microsecond speeds, vastly increasing the capabilities of adversaries. Mandia shares that since January, his team at Armadin has identified over 90 zero-day vulnerabilities in production environments belonging to Fortune 500 companies, illustrating how AI-driven attacks have become a significant threat. He underscores that the traditional methods of cybersecurity are inadequate to counteract this tsunami of AI-powered attacks, making autonomous defense systems essential.

Mandia describes the current state of AI-led attacks, noting that these models can operate via thousands of agents working collectively, which is unprecedented. He highlights how AI's ability to scan and probe multiple pathways simultaneously dwarfs human efforts in scale and speed, making attacks exponentially more effective. Mandia explains that AI can perform exploits based on structured code vulnerabilities faster and more comprehensively than humans. He warns that as access to GPU resources becomes easier and more anonymous, criminal AI attacks are likely to increase, creating a new battleground that favors attackers.

Mandia compares traditional nation-state cyber operations, which tend to be precise and focused, to AI-driven attacks, which resemble drone swarms—more noisy, broad, and destructive. He predicts that future conflicts will involve AI swarms that can overwhelm targets, forcing nations to reconsider their offensive doctrines. He also talks about attribution challenges, as AI-based attacks make it harder to identify who is behind them, whether nation-states or individual hackers, complicating response strategies.

Mandia highlights Armadin’s approach to cybersecurity: continuously testing networks with AI-driven hyper-attacks to discover exploitable risks before malicious actors do. The company maps networks rapidly using drone swarms, identifies changes, and attacks vulnerabilities in real time. Mandia stresses the need for autonomous response systems capable of reacting instantly to threats, as humans cannot keep pace with AI-operated attacks. He envisions a future where security controls are automated and proactive, reducing exposure windows and preventing breaches.

Mandia critiques traditional pen testing, arguing that it only uncovers known vulnerabilities and does not verify exploitable risks in real-world scenarios. In contrast, Armadin's AI-based approach actively exploits vulnerabilities in production systems, providing a more accurate assessment of true risk. He asserts that pen testing will become obsolete, replaced by continuous AI-driven probing that simulates real adversaries. Mandia emphasizes the importance of integrating offensive AI techniques into defensive strategies to stay ahead of malicious actors.

Mandia discusses how AI will reshape every layer of cybersecurity, necessitating autonomous prevention and response. He advocates for building security systems that can automatically block and respond to threats within seconds. Mandia notes that security teams must evolve from detect-and-respond to preventive, autonomous operations, and suggests that traditional detection methods will be phased out. He also stresses the importance of organizational discipline, rapid scaling, and clear processes to manage the chaos that comes with AI’s rapid pace, emphasizing that security excellence depends on making customers happy and differentiating through customer satisfaction.

Mandia analyzes incidents like the Hugging Face scenario, noting that underestimating AI models' capabilities is a common mistake. He talks about how companies often wish to slow AI development but are hindered by the race for innovation, which limits control. Mandia stresses that security must keep pace with AI’s rapid development, including logging and replaying AI actions for accountability. He advocates combining domain expertise with AI knowledge to secure agents and prevent rogue behaviors, illustrating his point with how Armadin uses layered validation and human oversight in its systems.