On 16 July, staff at the US online AI tool hub Hugging Face announced a “security incident disclosure”. Internal data had been accessed without authorisation. But it was the perpetrator, rather than the target, that caught the eye. Hugging Face said the breach “was different from anything we had handled before in one important way: it was driven, end to end, by an autonomous AI agent system”. Five days later, OpenAI, the company that introduced the world to AI when it released ChatGPT in 2022, announced that its programs were responsible for the hack.
The only technical account of the incident so far comes from the companies themselves. OpenAI reported that it was testing an AI “agent” – a system that can act on its own, within parameters – and its ability to exploit vulnerabilities in code. The agent found a way to escape from its “sandboxed testing environment” and access the internet, looking for answers. On 30 July, the AI company Anthropic’s model breached the systems of three other companies in a similar way.
These incidents are proof, if any more was needed, that the cybersecurity risks of AI models are real and growing. But this is not just the failure of a model or a lab. The entire international political community has failed. Cybersecurity has been high on the list of potential AI risks for years. OpenAI and Hugging Face patched up the incident between themselves, but such resolutions cannot be relied on. There should already be shared, established standards and protocols for incidents like this.
World leaders have wasted precious time treating AI as a zero-sum contest between Washington and Beijing, one with a finish line and a winner, in which cooperating with the other side looks like giving something away. Most AI experts agree that handling the technology’s risks requires global coordination. So why is there still no global policy infrastructure, particularly between the two biggest AI superpowers, the US and China?
There have been some useful efforts: bilateral dialogues, technical working groups, the odd joint statement. Several countries, including the UK, have established AI safety or security centres in an attempt to support the safety and security of the system. But isolated national efforts are not enough, and international cooperation has been obstructed by the pervasive and false view that AI is like the Cold War arms race.
Vladimir Putin may have come across as hostile and extreme when he said of AI in 2017: “Whoever becomes the leader in this sphere will become the ruler of the world.” But, by 2026, this view has become consensus. Politicians and technologists in the US, particularly, believe that they are in a race to achieve total supremacy in AI, and that there is some technical Rubicon that they must cross first, no matter the cost.
If AI is an arms race, and an existential one at that, then what will be permitted in the attempt to win it? The approach becomes: do not slow down, and do not help anyone else speed up. Safety standards or lessons learned are hoarded rather than shared in everybody’s interest.
But AI security is defined by mutual vulnerability: no state can be secure while others are exposed because the proliferation of AI is too great. Research by the AI and Geopolitics Project, of which I am director, at Cambridge’s Bennett School of Public Policy shows that it is possible for states to protect their national interest while cooperating on AI security in the global public interest. Everyone needs faster, clearer channels for disclosure when a system does something its operators didn’t intend. This is exactly the kind of narrow, technical ground on which countries that agree on almost nothing else can find common cause.
There is a precedent for choosing cooperation over competition, even between rivals. When the US and the Soviet Union were locked in the Cold War – a geopolitical rivalry many believed could end only in one side’s total victory – they still agreed the 1967 Outer Space Treaty, which kept weapons out of orbit and off the moon.
Washington pursued technological superiority, but did so alongside a rules-based approach. We remember today that America won the race to land on the moon; we largely forget the diplomacy that made the achievement one which the rest of the world could share rather than fear.
AI, too, is a frontier. Cybersecurity is an opportunity for practical cooperation, so long as we choose to treat recent AI security breaches as an argument for it, rather than as fresh ammunition for competition. The Space Race metaphor could divert us from a narrative of total domination towards one of shared discovery, without compromising national security imperatives. Forgetting the lessons of history allows Putin’s vision to dictate the trajectory of the future.
That is a political choice. We are not on a predetermined path. We arrived at this trajectory because of a series of decisions about how to behave towards this technology. We can choose to do it differently. The Space Race showed that self-interest and self-restraint are not incompatible, that competition and collaboration can coexist. If we forget that history lesson, the next AI security incident might not end so amicably.
Verity Harding is the author of AI Needs You: How We Can Change AI’s Future and Save Our Own (Princeton)
[Further reading: The Kemi-naissance is bullshit]
Content from our partners
Related
This article appears in the 12 Aug 2026 issue of the New Statesman, A Very British Scandal
Join the debate
Subscribe here to comment