DeepSeek banned and the AI cold war deepens

The United States government has banned DeepSeek AI from all federal devices as technology tensions with China continue to escalate, the BBC reported this week. The ban covers DeepSeek’s suite of AI applications, which attracted global attention in January 2026 when the Chinese company released a series of large-language models that performed comparably to leading American systems at a fraction of the reported training cost. Simultaneously, court filings in a federal case revealed that a former OpenAI employee accused of planting a tracking device inside the company’s San Francisco offices was allegedly operating with financial and logistical support from Saudi Arabia. The Al Jazeera report cited prosecutors as saying the accused, a former AI researcher, had received payments routed through a Saudi-linked entity. The two stories — Chinese AI banned from government, Saudi intelligence allegedly infiltrating the world’s most prominent AI safety company — arrived in the same news cycle and should be read together. They mark the moment that artificial intelligence definitively ceased to be a technology sector story and became a national security one.

The received wisdom

The standard framing treats the DeepSeek ban as a relatively straightforward data-security decision, analogous to the TikTok restrictions that have been litigated for several years. Chinese-developed applications, the argument runs, create potential vectors for data exfiltration and may contain backdoors accessible to the People’s Liberation Army or the Ministry of State Security. This is an established concern that applies to any technology stack with Chinese ownership — hardware, software, or infrastructure — and the DeepSeek ban is simply the latest application of a consistent principle. The OpenAI espionage case, under this framing, is a separate matter: a disturbing but isolated incident of a rogue employee allegedly recruited by a foreign intelligence service, of the kind that has affected sensitive American institutions for decades. The appropriate response is better security vetting, not a broader revision of how AI companies govern access and insider risk.

A different read

The two stories are linked by something more than timing, and the mainstream framing’s insistence on treating them as separate misses the most important signal. What both stories reveal is that the global AI competition has entered a phase in which the principal adversaries are no longer simply racing to build better models — they are now actively trying to understand, replicate, disrupt, and penetrate each other’s development processes. That is not a technology competition anymore. It is an intelligence competition, and it has been happening largely inside institutions — OpenAI, Anthropic, DeepMind, Google DeepMind — that were built for open scientific collaboration and have very limited experience managing the kind of insider threat environment that defence contractors and intelligence agencies have spent decades developing procedures for.

The DeepSeek story is more interesting than the ban. When DeepSeek released its models in January, the reaction in Silicon Valley was a mixture of admiration and alarm. The company had apparently achieved frontier-level performance using training techniques that reduced compute requirements dramatically — either because Chinese researchers had made genuine algorithmic breakthroughs, or because DeepSeek had access to model weights or training data that it had not disclosed, or because the cost claims themselves were misleading. The BBC reported that the ban is partly driven by concerns that DeepSeek applications may route user data to servers in China, where it would be accessible to the Chinese government. But the more sophisticated concern — discussed in security circles but rarely in mainstream coverage — is about what DeepSeek’s capabilities tell us about the state of Chinese AI development and about whether American export controls are creating the headroom that policymakers believe they are.

The Musk and Anthropic announcement of public share giveaways sits in this same landscape: AI companies are now significant enough that their governance structures — who owns them, who advises them, who infiltrates them — matters to national security in the way that Boeing’s governance matters to the Pentagon. The OpenAI espionage case should accelerate a conversation that AI companies have been reluctant to have: they are operating in a sector that multiple foreign intelligence services have identified as a primary target, they have historically hired globally from a talent pool that includes nationals of adversary states, and their internal culture of openness and collaboration is precisely the vulnerability that a sophisticated intelligence operation would exploit.

This is not a call for paranoia or xenophobia. It is a call for institutional maturity. The semiconductor industry learned this lesson over decades of corporate espionage cases. The pharmaceutical industry learned it with COVID-related research theft. AI is simply the latest domain where the value of the intellectual property has risen to the point that it attracts the full toolkit of state-sponsored intelligence operations. Anthropic’s CEO Dario Amodei has said publicly that the company considers itself to be working on one of the most consequential technologies in human history. If that claim is taken seriously, the security infrastructure needs to match the stakes.

The broader right-of-centre point is about the role of government in this space. Conservatives who are sceptical of technocratic overreach have generally been suspicious of the regulatory frameworks being proposed for AI — the European AI Act, the proposed American AI safety legislation, the various executive orders around model evaluation. That scepticism is legitimate when it targets poorly designed mandates that would impose compliance costs without security benefits. But the DeepSeek ban and the OpenAI espionage case are a reminder that there are genuine national security dimensions to AI governance that go beyond the progressive “alignment” discourse. Keeping the most capable AI systems out of adversary hands, and keeping adversary intelligence services out of the institutions developing those systems, is a conservative interest — one that should command as much attention from the right as the fiscal cost of regulatory compliance.

What to watch

  • Congressional hearings on AI security: the OpenAI espionage case will almost certainly generate subpoenas and testimony requests; watch whether the hearings focus on the specific incident or on the broader security posture of the AI sector.
  • DeepSeek’s response to the ban: the company has previously denied that it routes data to Chinese servers; a detailed technical rebuttal, or a significant silence, will both be informative.
  • Whether the ban extends to allied governments: if the UK, Australia, and Canada follow the US in restricting DeepSeek on government devices, it suggests a coordinated Five Eyes assessment rather than a unilateral American decision.
  • OpenAI’s internal security review: the company’s board and investors will be under pressure to commission an independent assessment; the results — or a refusal to release them — will signal how seriously the institution is taking the threat.

— J