AI Models Outpace Defenses, Shifting Focus to Cybersecurity Investment
Attempts to block AI models for cyber defense are futile; experts urge a strategic shift towards robust defensive measures and shared responsibility between industry and government.
2026 has emerged as the year when the long-feared potential of AI-powered cyberattacks is becoming a reality. New AI models now possess capabilities on par with skilled human hackers, marking a critical juncture for both AI development and cybersecurity policy. This transitional period is straining existing cybersecurity infrastructure, raising urgent questions about how to manage the risks introduced by AI before they overwhelm defenses.
Efforts to control access to AI models with potent cybersecurity applications, such as the U.S. government's temporary export controls on Anthropic's Mythos and Fable models, are proving to be short-lived solutions. History shows that as one generation of AI models advances, others quickly follow. OpenAI's GPT-5.5 and China's Z.ai GLM-5.2 are examples of models rapidly approaching or matching the capabilities of leading AI systems in cybersecurity applications. The global race to develop and release increasingly powerful AI models, often with open-weight architectures, makes centralized control nearly impossible. Anyone with sufficient computing resources can adapt these models for malicious purposes, rendering traditional access restrictions ineffective.
The only sustainable long-term strategy lies in significantly bolstering defensive capabilities. However, current defensive efforts are struggling to keep pace with the rapid advancements in AI. The federal government's reallocation of resources and authorities away from key agencies like CISA has created a void. This gap has been partially filled by AI companies themselves, which have initiated programs like Anthropic's Project Glasswing and OpenAI's Patch the Planet to secure critical infrastructure and open-source software.
While AI companies have incentives to invest in cybersecurity—ranging from public relations to securing their own supply chains—their motivations are not aligned with national security. Their primary focus remains on limiting liability and mitigating corporate blowback, rather than comprehensively securing national infrastructure or protecting citizens. The public commitments from companies like OpenAI and Anthropic to bolster U.S. cyber defense are valuable but represent only a partial solution to a much larger, systemic problem.
AI companies can leverage their powerful models to identify software vulnerabilities and even generate patches, a crucial capability. However, the more significant challenge lies in ensuring these patches are effective and can be deployed across critical systems without causing further disruption. This is particularly true for critical infrastructure, which often relies on aging, understaffed, and continuously operating systems that are inherently fragile.
AI companies share a responsibility for cyber defense, especially given the new threats their technologies enable. However, this responsibility is not theirs alone; it is a shared burden with other corporations and, crucially, the government. Owners and operators of critical infrastructure, government agencies, and private corporations all require a reliable source of information to assess evolving risks and identify effective mitigation strategies.
Traditionally, the federal government has served as the central clearinghouse for threat intelligence, gathering information from both public and private sectors and disseminating guidance. The government's role in responding to and recovering from cyberattacks should remain its purview, not be delegated to AI companies. The future will undoubtedly bring more cyberattacks, regardless of whether they are AI-powered.
To strengthen national defenses, leaders must prioritize measuring exposure to attacks, rigorously testing system resilience, and reducing recovery times. While AI companies can contribute to addressing the new threats they've helped create, they cannot replace the fundamental role of government in national cybersecurity strategy. The federal government's current approach appears reactive rather than proactive, highlighting the urgent need for a long-term cybersecurity strategy that moves beyond quick fixes like blocking individual model releases. The threat is evident to all; the critical question is whether there is the collective will to act decisively before it is too late.