U.S. and China Warn AI Could Trigger Accidental Nuclear War

WASHINGTON — Security researchers in the United States and China are warning that the rapid integration of artificial intelligence into military defense systems could make autonomous decision-making in strategic hardware an unprecedented risk for accidental nuclear war.

A joint policy report from the Brookings Institution in Washington and Fudan University in Shanghai calls on U.S. and Chinese leaders to establish binding international red lines. Those limits would prohibit artificial intelligence systems from independently executing cyberattacks against nuclear command networks or making launch decisions.

The report’s authors directly called on President Donald Trump and Chinese President Xi Jinping to formalize bilateral agreements preventing non-human systems from interfering with nuclear command, control, and communications (NC3) networks or critical strategic infrastructure.

The proposal comes amid heightened friction over the militarization of commercial AI technology and the authority that automated software should receive during national security crises. In February, AI developer Anthropic refused defense requests to lift internal safety guardrails restricting its Claude model from use in mass surveillance operations or fully autonomous lethal weapons platforms.

Nuclear command, control, and communications infrastructure includes complex networks that process satellite early-warning data, maintain continuous contact with strategic forces, and transmit firing orders to missile silos, strategic bombers, and ballistic missile submarines. U.S. assets include space-based infrared satellite constellations, ground radar installations, and specialized airborne command platforms such as the E-6B TACAMO aircraft.

Following the dispute with Anthropic, President Trump directed federal agencies to terminate technology contracts with the company. A federal judge overturned that action on August 28, ruling that the executive administration had engaged in unlawful retaliation against the firm.

Because nuclear engagement decisions may have to be carried out within minutes, experts warn that autonomous AI software in NC3 architectures could produce critical errors. An algorithmic misreading of sensor data, an automated reaction to a false alarm, or an unverified cyber-intrusive operation could trigger rapid escalation before human commanders have time to assess the threat.

The vulnerability of automated early-warning infrastructure has a historical precedent. In September 1983, a technological glitch in the Soviet Union’s Oko satellite system falsely indicated that five incoming U.S. intercontinental ballistic missiles had been launched.

Stanislav Petrov, the duty officer at the command center, recognized the alert as a false alarm and decided not to notify military leadership, preventing a potential retaliatory strike. Cold War commanders later restricted the role of automated systems, although the Soviet Union subsequently built “Perimetr,” known in the West as “Dead Hand.” It was a semi-automated backup system configured to launch retaliatory strikes if high-level communications links were severed.

Modern AI integration adds concerns about the speed and unpredictability of current machine-learning models. In February war-game simulations examining algorithmic strategic behavior, commercial foreign and domestic large language models, including systems developed by OpenAI, Google, and Anthropic, selected nuclear strikes in 95 percent of tested crisis scenarios.

Researchers cautioned that the simulation parameters may have encouraged aggressive moves, while noting the volatile tendencies of current models under operational stress. The Brookings-Fudan report proposes applying established human-control protocols for nuclear weapons directly to military cyber operations targeting strategic systems.

The recommended safeguards include dedicated military AI communication hotlines, a shared terminology defining effective human supervision, and verifiable measures to prevent unauthorized or automated execution.

“Presidents Donald Trump and Xi Jinping can lead in governing these dangers by agreeing that only humans, not AI, should make the decision to initiate a cyberattack against each other’s nuclear command, control, and communications systems (NC3) or critical infrastructure,” the authors wrote. They emphasized that “the governance of military AI is not a negotiation that can afford to wait—it’s a race against time.”

The legal dispute involving Anthropic also highlighted broader industry concerns about the speed of military AI implementation. Anthropic Chief Executive Officer Dario Amodei has publicly urged AI developers to deliberately slow the development of frontier model capabilities until robust safety controls are in place.

Amodei’s call for restraint was supported by OpenAI Chief Executive Officer Sam Altman and xAI founder Elon Musk. OpenAI later submitted inquiries to Congress seeking guidance on whether leading technology developers could coordinate voluntary safety limits without violating federal antitrust regulations.

U.S. defense policy already addresses human involvement in autonomous warfare through Department of Defense Directive 3000.09. First issued in 2012 and updated in January 2023, the directive requires all autonomous and semi-autonomous weapon platforms to allow commanders and operators to exercise “appropriate levels of human judgment over the use of force.”

Diplomatic discussions on AI governance have moved slowly. At the November 2023 summit in Woodside, California, former President Joe Biden and President Xi Jinping reached a general agreement affirming the need to maintain human oversight over decisions to employ nuclear weapons.

Official bilateral safety talks followed in Geneva, Switzerland, in May 2024, focusing on technical risks associated with advanced AI deployments. The report does not require Washington and Beijing to agree on broader domestic cybersecurity frameworks or infrastructure defense policies.

Instead, the Brookings-Fudan proposal is narrowly focused on preventing human error, software unpredictability, or unauthorized machine actions from initiating a strategic military conflict.