When the world's most circumspect intelligence community compresses the timeline for "frontier AI models overwhelming corporate defenses" from years into months, the race over how quickly AI tools can be put into defenders' hands has already outgrown the technical question itself.
The Five Eyes Warning: Months, Not Years
In June 2026, the intelligence agencies of the Five Eyes alliance (the United States, the United Kingdom, Canada, Australia, and New Zealand) jointly issued an unusual statement: AI models capable of launching cyberattacks strong enough to overwhelm the defenses of governments and enterprises may be only months, rather than years, away.
The force of this judgment lies in its source — the intelligence community most cautious about new technologies. Five Eyes is not a tech company's market-forecasting department, nor a trend report from an academic research institute; it is an alliance whose reason for existing is a low false-positive rate. When it says "months," cross-validation between signals intelligence and human intelligence stands behind the phrase.
When AI Plays Both Spear and Shield
The core argument of the Five Eyes statement is two-layered — AI lowers the barrier to entry for malicious cyberattacks: adversaries no longer need top-tier cybersecurity experts to mount high-intensity attacks; and at the same time, AI raises the speed and sophistication of attacks: automated chains of vulnerability discovery and exploitation have drastically compressed the time window.
This is a classic scenario of "AI accelerating the race between spear and shield." On the very same day, Five Eyes offered contradictory guidance: on one hand warning that defenders must "act immediately" to harden their networks, while on the other acknowledging that defense itself is also being made more efficient by AI — security operations teams that integrate AI tools can discover vulnerabilities earlier and respond to incidents faster. In other words, AI is making both attack and defense more effective; the question is which side wins the race against time.
The Self-Wounding Loop of Trust
The timing of this warning is worth noting. Less than two weeks earlier, the U.S. government had barred foreign nationals from accessing Anthropic's two newest models — Fable 5 and Mythos 5 — on national-security grounds. Anthropic was forced to take both models offline worldwide. A ban "intended to protect security" landed just days before a joint statement calling for global cooperation on defense.
An analysis published in Foreign Relations magazine laid bare the paradox behind this: America's AI export controls are manufacturing a trust deficit within the alliance system. At the G7 summit, Canadian Prime Minister Mark Carney warned that if model performance can be covertly throttled by developers, and if the U.S. government can sever access at any moment without warning, then allies will not adopt the American AI technology stack — and what today's global defense architecture needs most is precisely the sharing of AI cybersecurity tools.
The core ambition of "Pax Silica," the U.S. State Department's flagship initiative on AI and supply-chain security, is to build a U.S.-centric global AI ecosystem. But the combination of the "Anthropic ban + Five Eyes warning" is generating a contradictory signal: Washington asks allies to harden their defenses with American AI tools while simultaneously telling them those tools can be remotely disabled at any time. When allies are forced to choose between two forms of dependency — depending on American AI versus abandoning it — distrust itself becomes the largest security vulnerability.
The Exposure of Small and Medium Businesses
The Five Eyes agencies address "leaders of governments and enterprises," but comments from Olivia Shen, director of the Strategic Technology Program at the University of Sydney's United States Studies Centre, shifted the focus to a more vulnerable group — small and medium-sized businesses: "The sophisticated businesses, usually large enterprises, have already invested in cybersecurity, so they are better prepared. Those at greater risk are the small and medium enterprises that have been under-invested so far — they are essentially lambs to the slaughter."
When AI models can autonomously discover and exploit security vulnerabilities within months, the gap between enterprises that can procure the most advanced AI defense tools and enterprises that lack even a basic IT security team will change from a difference of quantity into a difference of quality — the latter will be directly exposed within range of automated attacks, with no buffer of human intervention.
The Tug-of-War Between Control and Openness
One key question raised by the Five Eyes warning remains unanswered: how will the United States strike a balance between "keeping frontier models out of adversaries' hands" and "getting AI security tools into as many defenders' hands as possible"?
The current mode of issuing case-by-case bans is producing three side effects:
An enforcement vacuum in capability control. The implementation details of the Anthropic ban remain vague: companies do not know how to prove that "users are not misusing the models," how to verify users' nationality, or how to comply with directives without disrupting legitimate business. When uncertainty becomes the primary output of policy, companies choose to wait and see rather than accelerate deployment.
A wasted window for defense. The limited cybersecurity testing capacity of frontier labs cannot cover most open-source projects and industrial control systems. Even if the United States can complete restrictions on public models within months, what is lost is the same span of mere months of defensive preparation time.
Centrifugal force in the alliance system. Carney's concerns at the G7 summit were not an isolated case. If Washington wants allies to adopt the American AI technology stack — and to give up Chinese alternatives — it must reassure them that the stack will not be remotely locked down one day.
From Early Warning to Intervention — The White House's Manual Control over Frontier Model Releases
In late June 2026, the White House intervened in the release of OpenAI's newest model, GPT-5.6 — the model will not be opened up as the company originally planned, but will instead be made available only to a handful of U.S. companies and organizations approved by the Trump administration. This was no gentle reminder but direct administrative intervention: in a leaked internal memo, OpenAI CEO Sam Altman acknowledged that a controlled release was not the company's "first choice," but that it chose to cooperate in order to build a sustainable framework of collaboration with the government.
This decision came at a moment when the U.S. AI regulatory environment was in a peculiar state: President Trump had previously signed an executive order on governing AI models with advanced hacking capabilities, but the order "remains in its early stages and has yet to set clear rules of conduct for companies." Meanwhile, Anthropic's newest model, Fable 5, was forced offline days after its public release over safety concerns, and its other most powerful model, Mythos 5, is open only to a few trusted research institutions. OpenAI itself likewise gave its most capable cyber model, GPT-5.5-Cyber, a restricted release.
Anthropic's Fable 5 pulled from public release → GPT-5.6 throttled by the White House → an executive-order framework that remains incomplete even as intervention becomes the norm. The three threads together reveal one fact: U.S. control over the most frontier AI models has shifted from "notifying" to "intervening" — yet the basis and the boundaries of that control are still hanging in the air.
Control Arrives, but the Rules Are Missing
In its statement, the White House said it would continue working with frontier AI labs to jointly develop responses, but one glaring contradiction exists at the operational level: the executive order asks companies to voluntarily submit models for review 30 days in advance, yet on the core questions — "which models need review" and "what are the evaluation criteria" — there has been no follow-up at all.
Dean Ball, a former AI policy adviser in the Trump administration, made a prescient judgment after the Fable 5 episode: "I suspect that until the U.S. government resolves the Fable incident it stumbled into, public releases of new products across the entire American AI industry will effectively come to a standstill." Shortly afterward, he joined OpenAI. That means the people who best understand the logic of regulation are migrating to the regulated side.
Regulatory Opacity Is Itself a Security Vulnerability
The White House's current operating mode can be summed up in one formula: case-by-case intervention, lagging rules. Each intervention resolves the risk of the moment, but every ambiguous boundary plants the setup for the next conflict. What AI companies face is not clear compliance requirements but a permitting logic that is "constantly changing and always kept confidential" — as Ball put it, the government "can only grope its way toward rules in real time, ad hoc, reacting to incidents as they happen."
This pattern has three deep implications for U.S. AI security strategy:
Loss of rhythm. When companies cannot predict the administrative approval cycle for releases, R&D planning can only shift from "when it is ready" to "when it is allowed." For an industry whose core competitiveness is iteration speed, an uncertain approval cycle is itself a form of efficiency loss.
Distorted flows of innovation. If the release pathways of the most frontier models are locked down by case-by-case management, resources will naturally flow toward two kinds of scenarios: weaker general-purpose models that require no review, and application-layer deployments inside closed internal systems. Neither is conducive to building shared AI security capabilities in an open defensive environment.
An unbalanced global competitive landscape. When U.S. and Chinese AI companies develop a systematic gap in product-release rhythm, the side waiting for approval automatically cedes the market window. The June executive order was meant to be a safety measure, but its actual effect — especially when Chinese AI companies face no parallel constraints — may conversely accelerate the erosion of U.S. competitiveness at the AI application layer.
From a threat warning measured in "months" to case-by-case regulatory intervention, within the span of a single month U.S. AI security policy has upgraded simultaneously at both the cognitive and the operational level. Whether these two upgrades point in the same direction remains an open question: defense needs firm rules and a predictable framework of cooperation, while what current regulation is outputting is uncertainty and case-by-case boundary-testing.
The Five Eyes "months" warning exposes a fundamental contradiction in cybersecurity in the AI era: defense requires sharing and coordination, but America's current reactive regulation is creating an environment of trust deficit. When the evolution of the "spear" outpaces the coordination capacity of the "shield," the problem of technical security is transforming into a problem of trust — and the latter has never been solvable by unilateral regulatory orders.
China's Counter-Cyclical AI Expansion — The Market Window Created by Regulatory Wobble
While U.S. AI policy burns through its own credibility and time window in a cycle of "loosen — restrict — loosen again," Chinese AI companies are responding not by waiting but by accelerating delivery. From late June to early July 2026, at least two Chinese companies announced security-grade AI systems capable of competing with Anthropic's Mythos and OpenAI's top models — at a fraction of the price of their U.S. counterparts.
360 Security Technology released two AI models that the company claims offer functionality comparable to Anthropic's Mythos. Reuters reported that these tools are designed to substantially enhance vulnerability discovery and automate responses to cyberattacks. 360's approach directly targets Mythos's strength — vulnerability discovery — distilling twenty years of offensive-defensive security experience into model capability rather than relying on the emergent abilities of general-purpose large models.
Zhipu AI (Z.ai) released GLM-5.2 at roughly one-sixth the price of leading U.S. models. More critically, GLM-5.2 is an open-source model — users can download and modify it directly, which means the barrier to removing its built-in safety measures is extremely low. Independent assessments by cybersecurity company Semgrep and visual-investigation platform Graphistry indicate that GLM-5.2's vulnerability-discovery capability is "on par with leading U.S. models."
Semgrep founder Isaac Evans assessed that while GLM-5.2 and Mythos "belong to different classes," the model's capability still represents "a huge leap compared with other models." Graphistry researchers speculated that GLM-5.2's capabilities may derive in part from "illicit distillation" — that is, knowledge distillation performed on the API outputs of OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8.
Regardless of the legality of the distillation, one structural fact will not change: when U.S. policy wobble causes frontier model releases to come and go in fits and starts, what Chinese open-source AI models fill is a market vacuum — not a technology vacuum. Security experts and cyber hawks on Capitol Hill worry that America's current release restrictions on its own models are, paradoxically, accelerating the iteration pace of Chinese models.
The warning from Representative Andrew Garbarino, chair of the House Homeland Security Committee, embodies this anxiety: "China may be only months, or even weeks, away from achieving frontier AI capabilities on par with the United States." Matt Perl, former director for emerging technology at the National Security Council during the Biden administration, offered a colder assessment: "The United States and China are indeed engaged in an arms race to develop and deploy the most advanced AI technologies. Just as during the Cold War, both sides see this as an existential contest."
What this means: the wobble in U.S. AI policy has created a discontinuous time window, and Chinese AI companies have used the gaps of regulatory uncertainty to complete model deliveries and open-source diffusion. In this process, what the United States has lost is not only time — but also the confidence of allies and partners in the reliability of its AI supply chain. When "American technology control" becomes "American technology stall," the trust deficit spreads from the security domain into the industrial one.