AI arms race heats up: US models power China’s defense, while “open” rules fracture

On July 31, 2026, multiple outlets converged on a single, destabilizing theme: frontier AI is moving from research labs into defense and autonomous cyber operations. A DefenseNews report says Chinese military researchers used outputs from leading U.S. AI models—specifically OpenAI and Anthropic—to train domestic systems aimed at advancing defense capabilities. Separately, The Hacker News (via Palo Alto Networks Unit 42) describes a Chinese-speaking threat actor that used DeepSeek through the open-source Hermes Agent framework, receiving an initial Telegram instruction and then autonomously scanning the internet for exposed systems and selecting public exploits. Meanwhile, a National Interest piece frames the strategic vulnerability of America’s AI buildout, arguing that U.S. trade policy is poised to target the hardware supply chain that AI expansion depends on. Strategically, the cluster highlights a widening gap between public “AI safety” narratives and the operational reality of military and intelligence adoption. If Chinese defense researchers can legally or semi-legally leverage U.S. model outputs to accelerate domestic capability, the effective control point shifts from model training to downstream integration, evaluation, and deployment—areas that are harder to police. The open-source dispute—Anthropic and OpenAI clashing with parts of the tech industry over whether Chinese-origin open models should be freely available or restricted—signals that governance is becoming fragmented across the ecosystem, not centralized in governments. The likely winners are actors that can move fastest from model access to systemization, while the losers are those relying on slower, compliance-driven diffusion of capabilities. Market implications cut across semiconductors, cloud infrastructure, and cybersecurity spend. If U.S. trade policy targets AI-critical hardware, the near-term direction is toward higher input costs and tighter availability for compute-intensive deployments, pressuring AI infrastructure budgets and potentially lifting risk premia for suppliers tied to constrained export categories. Cybersecurity firms and platforms—especially those selling agent-detection, exploit mitigation, and threat intelligence—stand to benefit as autonomous attack workflows become more credible and scalable, with Palo Alto Networks’ Unit 42 findings reinforcing demand for defensive automation. Currency and rates are not directly cited, but the macro channel is clear: supply-chain friction can translate into higher capex for data centers and faster rotation into “security-by-design” tooling, affecting equities linked to semis (e.g., NVDA/AMD exposure) and enterprise security (e.g., PANW exposure) through sentiment and earnings expectations. What to watch next is whether policy and industry governance converge into enforceable restrictions rather than competing principles. Key indicators include: any U.S. export-control expansions tied to AI hardware, new guidance on model access or licensing for frontier systems, and measurable changes in how quickly Chinese-linked threat actors operationalize agent frameworks like Hermes Agent. On the cyber side, monitor for increased use of messaging-based command-and-control (Telegram) paired with autonomous exploit selection, plus faster weaponization cycles from initial instruction to target compromise. The escalation trigger is a sustained pattern of autonomous attacks that cause material outages or data loss, which would likely force governments to tighten both cyber and AI governance; de-escalation would look like clearer boundaries on dual-use model distribution and more consistent industry compliance mechanisms.

Geopolitical Implications

• 01

  Model access and output reuse are becoming a practical pathway for military capability transfer, reducing the effectiveness of controls focused only on training data or frontier model weights.

• 02

  Industry fragmentation over “open” versus “restricted” model availability may undermine coordinated safety governance and increase strategic uncertainty for governments and markets.

• 03

  Hardware export and trade restrictions risk backfiring by incentivizing alternative supply chains and accelerating indigenous AI buildouts elsewhere.

• 04

  Autonomous cyber operations tied to AI agents can blur lines between espionage and disruption, raising the probability of tit-for-tat responses and escalation-by-incident.

Key Signals

• —New U.S. export-control or trade-policy measures specifically targeting AI hardware components and compute supply chains.
• —Any formal guidance or licensing changes by OpenAI/Anthropic regarding access to models or outputs for defense-adjacent use cases.
• —Observed increase in autonomous agent frameworks in real-world intrusions, especially those using messaging-based command triggers.
• —Public-private coordination announcements on AI safety that include enforceable restrictions rather than voluntary norms.