AI export controls that create, rather than close, security gaps
AI export controls on Anthropic’s advanced cybersecurity models are government rules that first blocked, then selectively restored, cross-border access to the company’s Mythos and Fable systems, reshaping which organisations and countries can use frontier tools and pushing rivals to build independent alternatives.
The core story is not that the US imposed AI export restrictions; it is that those controls have ended up rearranging, not preserving, technological advantage. On June 12, the Commerce Department ordered Anthropic to suspend access to its latest models for all foreign nationals and to disable Mythos 5 and Fable 5 entirely, citing national security fears about their ability to identify thousands of flaws in critical software. Two weeks later, Commerce Secretary Howard Lutnick reversed course, first allowing Claude Mythos 5 back for more than 100 domestic “trusted partners” and then lifting restrictions on all of Anthropic’s models so they no longer require export licenses. That whiplash did not freeze the global field; it convinced others that relying on US AI policy is a strategic risk.
Picking domestic winners while pushing foreign rivals to build their own
The partial reopening of Anthropic model access was not a return to open markets; it was a selective distribution of power. Lutnick’s decision to redeploy Claude Mythos 5 to over 100 domestic companies and federal agencies meant the US government explicitly anointed a club of insiders who would keep a critical offensive–defensive security edge that others were denied. In parallel, the same day saw another leading lab limit rollout of its latest GPT-5.6 models to a small set of government-approved partners, reinforcing that this gatekeeping model is becoming the practical norm in US AI policy.
This sits awkwardly with the administration’s earlier executive order promising a light-touch regime where model makers could voluntarily submit advanced systems for review up to 30 days before release. The sequence sends one loud signal abroad: access to top US models is contingent on political favour and can vanish overnight. For security-focused developers outside this circle, building their own systems is no longer optional; dependence is a liability. Export controls meant to concentrate capability in friendly hands instead motivate foreign capitals and companies to treat independent AI capacity as a strategic asset.
360 Security’s Mythos rival shows why one-way transparency is unsustainable
The most concrete response to Anthropic’s lockdown came from 360 Security Technology, which announced an AI suite it claims matches Mythos. Its flagship tool, Tulongfeng, is designed to automatically discover software vulnerabilities, while companion system Yitianzhen automates cyber defence and incident response. 360 says Tulongfeng has already identified over 3,400 vulnerabilities, even though founder Zhou Hongyi concedes domestic base models still trail American equivalents by 20–30%. That gap has not stopped deployment; it has changed tactics.
Zhou argues that allowing US entities to enjoy a monopoly on systems like Mythos would trap others in “one-way transparency,” where their infrastructure is visible to foreign scanners they cannot match. He calls these AI systems a “cyber nuclear weapon” that cannot “be held only by others” and says his company is pursuing an “agent” route—combining mid-tier models with deep security expertise and local vulnerability databases—to close the gap without the same chip and compute stockpiles. In other words, restricting Anthropic’s tools did not keep the capability scarce; it legitimised aggressive catch-up as an urgent national-security project.
From covert steganography to quieter de-escalation
While governments were tightening AI export restrictions, Anthropic was running its own quiet control experiment inside its coding tools. Earlier this year, the company embedded hidden steganographic markers in Claude Code to detect proxy routing, account resellers and, notably, hostnames associated with known Chinese AI labs and gateways. The system altered the internal prompt using invisible Unicode markers and hid a domain list behind XOR and base64 encoding, effectively watermarking interactions to flag potential model distillation or unauthorised resale.
Engineer Thariq Shihipar described this as a March experiment to counter distillation and account abuse, adding that stronger mitigations now exist and confirming a fix to remove the hidden code by July 1. This rollback coincides with the Commerce Department lifting all remaining export and in-country transfer licenses for Claude Mythos and Fable models, which officials said followed joint work with Anthropic to address risks. The timing matters: as Washington steps down from its most aggressive posture and the company strips out covert competitor-detection code, both sides are signalling a desire to reduce open confrontation even while they retain new, less visible guardrails.
Geopolitical AI competition is now about ecosystems, not single models
The past month’s whiplash is a preview of how geopolitical AI competition will play out: through ecosystem choices, not one-off model bans. Initial US export controls cut off foreign nationals from Mythos and Fable on security grounds, a stance reinforced by a White House executive order expressing intent to protect domestic AI from foreign adversaries. The confluence of security fears and rapid model advances pushed officials to act, but those moves also told allies and rivals alike that their access can be collateral damage in domestic political fights.
The result is predictable. Security firms like 360 are racing to build autonomous alternatives, treating AI as “the new frontier of strategic deterrence.” Export controls that once aimed to keep dangerous capabilities at home are instead fragmenting the AI landscape into competing blocs of models, chips and data. The Trump administration’s later decision to lift restrictions on all Anthropic models and “pave the way” for business as usual does not undo that structural shift. Leaders now face a stark choice: either build shared frameworks that make cross-border AI security cooperation possible, or accept a future where every major power treats frontier models as strategic weapons and builds them in parallel, regardless of how many controls Washington writes on paper.
