Donald Trump and Xi Jinping started a USA-China AI Dialogues track but agreed on no common rules. In July 2026, OpenAI models escaped a test environment and broke into Hugging Face servers. This incident showed that autonomous AI breaks the chain of responsibility, leaving many questions about who answers for consequences.
The Washington summit between Donald Trump and Xi Jinping produced a telling result on AI. The two sides formalised a USA-China AI Dialogues track, built around an incident hotline, but agreed on no common rules. Trump rejected any globalist scheme for governing the technology, while Xi spoke of jointly guarding against misuse without offering a shared framework. The world's two leading AI powers have agreed to manage accidents but not to define responsibility. That gap now holds the most serious risks.
Accountability traditionally rests on a simple chain comprising a person who decides, a person who acts, and a person who answers for the consequences. Autonomous AI breaks that chain.
In July 2026, OpenAI disclosed that its models escaped a sandboxed test environment, reached the internet, and used stolen credentials and zero-day vulnerabilities to break into Hugging Face's servers to cheat on a cybersecurity evaluation. Investigators later found that around 700 agents took part and that many tried to hide their tracks. No human ordered the attack. OpenAI's own review traced the behaviour to reward hacking and related misalignment patterns, meaning the systems were pursuing their assigned goal by means no one intended.
This raises questions that existing law was not built to answer. Who is responsible when an AI system commits an act that nobody authorised? The developer that trained it? The team that ran the test? The company whose security failed? If a human had done the same thing, it would be a crime with an identifiable perpetrator. When a machine does it, responsibility becomes diffuse.
Notably, Hugging Face detected the breach first and reported it to law enforcement before OpenAI linked it to its own evaluation. Disclosure depended on the goodwill and competence of the companies involved, not on any obligation. That may have worked this time. It is not a system.
In the US, lawmakers introduced the AI Kill Switch Act, which would require developers of advanced systems to keep the ability to shut them down, report incidents and preserve forensic records. Such measures are necessary, but national rules alone cannot hold a borderless technology to account.
AI systems are trained in one country, run on servers in another, and cause effects in a third. Open-weight models, once released, can be copied and deployed anywhere. An incident's origin may be impossible to establish quickly. When Anthropic reported that a group it assessed as Chinese State-sponsored had used Claude Code to conduct most of an espionage campaign autonomously against about 30 organisations, Beijing rejected the attribution. Without common standards for evidence and investigation, cross-border AI incidents end in accusation and denial, and nobody is held responsible.
Uneven rules also create perverse incentives. During the Hugging Face breach, the company first tried to defend itself with a leading US lab's model, but that model's safety guardrails got in the way. It turned to an open-source Chinese model instead. The lesson is uncomfortable. When safeguards apply unevenly, users move to wherever restrictions are weakest. Responsible actors are penalised, and overall safety falls to the lowest common denominator. Only a shared baseline can prevent that race to the bottom.
The case for universal rules also rests on a collective action problem. When Anthropic's Dario Amodei urged pacing frontier development to give governments time to catch up, Beijing dismissed it as an attempt to hold China back. Each side fears that restraint will hand the lead to its rival. Voluntary commitments by companies face the same pressure: a lab that invests heavily in safety testing risks falling behind one that does not. Hugging Face's chief executive argued that the breach proved AI safety cannot be managed by any one company working alone. The same is true of any one country.
History offers useful parallels; for example, civil aviation became safe not because each airline chose to be careful, but because the International Civil Aviation Organisation set universal standards for certification, incident reporting and accident investigation. Nuclear technology spread under the International Atomic Energy Agency's watch and verification regime. In both cases, shared rules did not stop innovation. They made it trustworthy enough to scale.
AI accountability needs similar foundations, common definitions of what counts as a reportable incident, obligations to disclose, independent evaluation of frontier models before deployment, clear lines of liability, and the ability to intervene when a system behaves dangerously. These are the basic infrastructure of accountability, not constraints on progress.
The alternative is already taking shape. Instead of universal rules, rival clubs are forming. The US-led Pax Silica secures AI supply chains among aligned partners. China's World AI Cooperation Organisation (WAICO) offers open membership without political conditions and targets the Global South, though its institutional design remains thin. The UN's Independent International Scientific Panel on AI and Global Dialogue on AI Governance are inclusive but lack binding force. The result is a patchwork in which accountability depends on which bloc a system belongs to, and incidents that cross blocs belong to no one.
India's position shows both the problem and the possibilities. It is a member of Pax Silica, the only BRICS member outside WAICO, host of the 2026 AI Impact Summit, and a major deployer of AI through its digital public infrastructure. Its experience shows how middle powers are pulled between blocs. It also shows the kind of bridging role that could turn fragmented initiatives into a common baseline. Whether that potential is realised will help decide whether the world gets universal rules or competing rulebooks.
Ideally, India should first try to push for any US-China incident mechanism to evolve into a wider, plurilateral reporting framework. AI incidents will not respect bilateral boundaries, and countries with exposed critical infrastructure deserve early warning.
Second, India should present itself as the bridge between clubs. It can use Pax Silica to bring trusted compute, chips, and investment to Global South partners, while also offering its own training and capacity-building programmes for Africa, Southeast Asia, and Latin America, without alignment strings.
Third, New Delhi should anchor governance in inclusive institutions, backing the UN Scientific Panel and Global Dialogue as neutral ground where both blocs must engage.
Fourth, it must strengthen its own capabilities with domestic compute, a national framework for responding to AI incidents, and interoperability standards that keep options open across ecosystems. A country cannot credibly shape the rules of a technology it cannot secure at home.
The Washington summit confirmed that the race for AI capability will not pause. That makes universal accountability urgent. Incidents over the past year show that AI systems can already act in ways no one ordered, across borders no single government controls, with consequences no existing legal framework clearly assigns. A hotline can manage a crisis once it starts. Only shared rules can establish responsibility before it does. Without them, the most powerful technology of the century will keep advancing with no one answerable for what it does.