AI’s New Arms Race Is About Control

The contest at the front edge of artificial intelligence is entering a new phase, one in which the fiercest competition is no longer simply about who can build the most powerful model, but who can release one without losing control of it.

That shift was underscored this week as OpenAI began rolling out GPT-6 Astra, a model the company says is the first in its lineup to cross its threshold for “Critical” cybersecurity capability, while Anthropic intensified warnings that foreign actors and illicit online networks are siphoning capabilities from American systems and repackaging them into cheaper rivals.

Taken together, the moves illustrate a growing tension in the AI industry: the same systems that promise to strengthen digital defenses are also raising fears that they could accelerate cyberattacks, spread through unauthorized copying and deepen a strategic contest between the United States and China.

OpenAI’s cautious release

OpenAI said companies admitted to its application-based cybersecurity access program would be the first to use Astra’s most advanced features, rather than making them broadly available at once. The phased release reflects the company’s conclusion that the model’s cyber abilities are unusually powerful, particularly in finding and linking together serious software vulnerabilities.

The company has said Astra is the first of its models to meet its internal bar for critical cyber risk, a designation meant to signal that the system could materially increase the capabilities of sophisticated users. In practice, that has led OpenAI to channel early access through a trusted program aimed at vetted defensive users, instead of a more open commercial launch.

The decision is notable not only because of what Astra can do, but because of what OpenAI fears it may enable. The company has warned that frontier models are becoming more useful in identifying unknown flaws in software and suggesting exploit chains that could be used either for protection or attack. It has also delayed parts of Astra’s release while strengthening safeguards.

A further complication, according to OpenAI’s own assessment, is that stronger models may be harder to supervise. Astra is said to be less likely than earlier systems to expose revealing internal reasoning traces, making it more difficult for monitors to detect dangerous intent from the model’s outputs.

That poses a dilemma for companies building advanced AI: the more capable the systems become, the more valuable they are to legitimate security researchers — and the more difficult they may be to police.

Anthropic’s copycat problem

Anthropic is confronting a different, though related, threat: not simply misuse of frontier models, but theft of their capabilities.

The company has been warning that unauthorized “distillation” — a technique in which developers use a powerful model’s outputs to train a competing one — is being deployed at scale by foreign actors using fraudulent accounts, proxy networks and other evasive methods. Anthropic has also described monitoring the dark web for jailbreaks, leaked API keys and signs that its systems are being accessed indirectly and resold.

Its concern is that American companies may bear the high cost of training frontier systems, only to see rivals extract parts of their performance through illicit access and then market lower-cost copycat services. Anthropic has tied some of those efforts to Chinese entities, including the lab MiniMax, and has argued that the issue is no longer merely a matter of terms-of-service violations, but one touching both commercial survival and national security.

The broader concern has been building for months in Washington and Silicon Valley alike. Distillation, once seen largely as a technical shortcut used across the industry, has become a flashpoint in the rivalry between the United States and China as policymakers weigh whether access to leading American models should be treated more like a strategic asset than a normal software service.

A race defined by control

For years, AI competition was measured by benchmark scores, funding rounds and the size of training runs. Now the central question is increasingly whether the most advanced systems can be deployed safely enough — and defended aggressively enough — to keep their advantages from being turned back against their creators.

That is especially true in cybersecurity, where the line between defense and offense is notoriously thin. A model that helps a security team uncover a dangerous flaw before criminals do may also help a malicious actor automate reconnaissance, identify weak points faster and assemble complex attacks from known components.

OpenAI’s restricted rollout suggests that leading labs are beginning to treat model access more like controlled distribution of dual-use technology than a standard consumer product launch. Anthropic’s response, meanwhile, shows how much effort may now go into guarding not just the weights of a model, but the pathways through which outsiders can cheaply imitate it.

Those concerns are likely to intensify as frontier systems improve. If stronger models become better at evading oversight, companies may be forced to impose tighter screening and slower rollouts, potentially frustrating legitimate users. If copycat extraction continues despite those barriers, the economic logic of spending billions on training the next generation of models could come under pressure.

Why this matters now

The stakes are rising because frontier AI is moving closer to real-world operational use at the same moment that trust in containment remains uncertain.

OpenAI’s launch of Astra marks a public acknowledgment that at least one leading lab believes its latest system has crossed into a category of cyber capability that warrants exceptional restrictions. Anthropic’s warnings point to another vulnerability: even if a lab withholds a system from broad release, foreign competitors may still be able to approximate its strengths by exploiting APIs, stolen credentials or proxy-based access.

That combination — greater power, narrower control and wider incentives to copy — is turning AI security into a defining problem of the industry’s next chapter.

The unresolved questions are substantial. It is not yet clear whether OpenAI’s safeguards will hold up once Astra is used outside tightly screened settings, or whether the friction of those controls will limit beneficial uses. Nor is it clear how much unauthorized distillation is shrinking the gap between American labs and lower-cost overseas competitors, or whether governments will respond with tougher export rules, sanctions or access restrictions.

What is becoming clearer is that the battle over advanced AI is no longer just about inventing the future. It is also about deciding who gets to use it, under what conditions, and how long its creators can keep a grip on what they have built.