AI and the New Global Security Landscape

What It Covers

Gordon Goldstein, adjunct senior fellow at the Council on Foreign Relations, joins Kevin Ball to examine how frontier AI models — particularly Anthropic's Claude Mythos — have created documented capabilities for zero-day cyberweapons, novel bioweapon synthesis, and autonomous battlefield targeting, forcing ad hoc US government oversight mechanisms into existence as of mid-2026.

Key Questions Answered

• •Zero-Day Machine Threshold: Claude Mythos represents a capability discontinuity, not a gradual progression. It identifies zero-days across 10,000 software systems simultaneously, exploits multiple vulnerabilities in parallel, generates original exploits that never previously existed, and remains undetected indefinitely. Cybersecurity experts describe this as previously science-fiction-level capability. Critically, 90% replication of this capability by open-weight Chinese models still constitutes an extraordinarily powerful offensive weapon.
• •AI Bioweapon Upskilling — Third-Party Validated: Independent red team tests confirm AI models provide step-by-step bioweapon synthesis guidance without requiring domain expertise. A 2025 red team test showed a model detailing how to weaponize genetic material and disperse it via weather balloon over a US city. A 2023 study prompted an AI to reward toxicity, generating 40,000 candidate chemical warfare agents — including novel molecules potentially deadlier than VX — within six hours.
• •Self-Regulation Is the Only Mechanism Currently in Place: No national legislation and no international treaty specifically governs AI-enabled bioweapon or cyberweapon creation as of July 2026. Frontier labs conduct internal red team tests and grade their own homework. The US government's oversight body, CASEY within the Commerce Department, lacks clear structure and rigorous analytical capacity. Congressman Josh Gottheimer has introduced mandatory NSA-led security review legislation, but it remains unpassed.
• •Open-Weight Chinese Models Circumvent US Oversight Entirely: Even a fully effective US regulatory framework cannot address threats from Chinese open-weight models, which are released publicly and replicate frontier capabilities within three to nine months. China's open-model architecture accelerates innovation and propagates security risks simultaneously. The Trump-Xi summit in spring 2026 identified AI security cooperation as a bilateral agenda item but produced no concrete mechanisms or commitments.
• •Autonomous Drone Warfare Erodes Human Decision-Making Control: Ukraine's conflict has shifted to primarily drone-on-drone warfare, where Russian soldiers sent to capture territory survive an average of 20–35 minutes. As drone swarms require instantaneous autonomous coordination that human operators cannot provide at speed, battlefield decision authority devolves to AI systems. Hypersonic weapons — capable of striking targets across continents in minutes — further collapse human response time, threatening the survivable second-strike foundation of nuclear deterrence.
• •Commercial Actors, Not States, Control the Technology: Unlike the nuclear era — where the US and USSR held near-monopoly state control over weapons technology — frontier AI capabilities reside almost entirely within private companies locked in ruthless commercial competition. Proposals for an IAEA-equivalent AI standards body have been floated by Demis Hassabis and Sam Altman, but no coordinating mechanism exists. The US-China joint understanding against AI nuclear control is informal and uncodified in any arms control treaty.