American frontier models keep pulling ahead. Chinese officials and analysts watch with mounting alarm. Last month in Singapore a representative from a Chinese think tank pressed Anthropic executives to reverse course and grant Beijing access to the company’s latest system. The answer was no.
The episode, first reported by The New York Times, captures a sharpening contest. Latest releases from Anthropic and OpenAI have widened the technology gap. They have also sharpened fears inside China that American advances could translate into lasting strategic advantage.
Anthropic’s model, called Mythos, can scan codebases, spot flaws in major operating systems and browsers, and suggest fixes at speeds that outpace human teams. The company limited its initial rollout to roughly 40 trusted organizations under a program dubbed Project Glasswing. Defense contractors, select cybersecurity firms and a handful of government partners gained early looks. Most banks, utilities and foreign governments did not.
OpenAI responded in kind. Days ago the company introduced GPT-5.5-Cyber, a specialized variant focused on vulnerability detection. Like Mythos it will reach only a narrow set of U.S.-based defenders at first. Cisco, CrowdStrike and Cloudflare sit on the approved list. The pattern is clear. Both labs treat these systems as sensitive instruments best kept from broad circulation.
Chinese analysts view the moves as more than commercial caution. They see a deliberate effort to deny them the tools that could accelerate their own progress. Matt Sheehan, senior fellow at the Carnegie Endowment for International Peace, told The New York Times that the episode marks a second wake-up call after ChatGPT’s debut years earlier. Export controls on advanced chips have already slowed Beijing’s hardware buildup. Now software breakthroughs threaten to stretch the timeline further.
Yet China refuses to sit idle. Its domestic market for AI-driven cybersecurity is forecast to reach $8.7 billion by 2030, a 37-fold jump from 2025 levels, according to research cited by the South China Morning Post. Local labs race to replicate the offensive and defensive capabilities on display in Mythos and GPT-5.5-Cyber. Some have gone quiet on public model updates, a signal that work has shifted behind closed doors, Germany’s top cybersecurity official warned lawmakers last week.
Claudia Plattner, president of Germany’s Federal Office for Information Security, told the Bundestag’s digital committee that providers including Alibaba appear close to matching the superhacking potential of Anthropic’s offering. Her assessment, relayed by Politico, underscores growing European worry that the technology will not remain an American preserve for long.
The stakes extend beyond national pride. Mythos and its peers can discover zero-day vulnerabilities faster than traditional scanning tools. They can also generate working exploits. In the wrong hands that power could paralyze critical infrastructure within hours. Google recently blocked a wave of attacks in which hackers directed an AI system to locate and weaponize flaws. The company stopped short of naming the model but noted it was neither its own Gemini nor Anthropic’s Mythos. The incident, covered by Fortune, illustrated how quickly adversaries adapt.
Even controlled release carries risks. Shortly after Anthropic announced Mythos a small Discord group reportedly gained unauthorized access by guessing internal paths and exploiting knowledge from a prior contractor breach. The company launched an investigation. Bloomberg first broke the story. The episode fueled skepticism about whether any lab can truly wall off such potent technology.
Rob Bair, head of cyber policy at Anthropic, defended the measured approach during a recent panel. He argued that giving network defenders a head start before broader exposure protects more than it endangers. His comments, quoted in Politico, reflect the tightrope both firms walk. Release too slowly and allies fall behind. Release too widely and attackers gain the upper hand.
Washington has taken notice. The Trump administration initially approached AI regulation with skepticism. Mythos changed the tone. Officials moved quickly to formalize review processes with Google, Microsoft and xAI. Earlier pacts with Anthropic and OpenAI gained fresh relevance. A Commerce Department announcement detailing the deals briefly appeared online then vanished. The confusion, reported by Fortune, highlighted internal debates about how much oversight these systems require.
Some voices inside the administration see Mythos as proof of American primacy. Treasury Secretary Scott Bessent reportedly called the model a breakthrough that could cement the U.S. lead over China. Others worry that superiority brings no immunity. The interconnected nature of global software means a flaw found in one jurisdiction can be turned against another within minutes.
Rest of World examined the widening divide. It noted that central banks and governments outside the trusted circle remain exposed while they wait for vendors to patch systems. Criminal groups already use earlier AI tools to craft phishing campaigns, clone voices and hunt weaknesses. A North Korean operation reportedly employed models from OpenAI and Cursor to steal millions in cryptocurrency. The article, published by Rest of World, painted a picture of uneven vulnerability that favors those with early access.
Beijing’s outreach in Singapore fits a broader pattern. Track-two channels have become testing grounds for quiet requests. When direct appeals fail, Chinese entities explore workarounds. Some labs train on synthetic data or older open-source foundations. Others probe for leaks. The Discord incident fed suspicions that determined actors could obtain the model through indirect routes. One expert told Fortune that if a random online group could reach it, state actors likely already had.
Still, the gap persists. DeepSeek, a Chinese lab watched closely in Washington, released new models last week claiming performance only marginally behind OpenAI and Anthropic on certain benchmarks. The announcement drew attention but also fresh accusations of technology theft, which Beijing denies. Coverage in Kyodo News described China as a vast testing ground for mass AI adoption even as its frontier models lag in raw capability.
The comparison to Cold War nuclear competition surfaces repeatedly among security analysts. Henry Kissinger co-authored an essay in Foreign Affairs that framed AI as the next domain requiring arms-control thinking. The analogy has limits. Software copies far more easily than fissile material. Yet the speed of iteration and the dual-use nature of these systems create parallel dilemmas. A single model can strengthen defense and enable offense at the same time.
U.S. policy reflects that tension. Export controls on chips continue. Reviews of powerful models before public release expand. At the same time the administration signals reluctance to slow domestic innovation. President Trump’s team has emphasized beating China as an existential priority. The recent overture from Beijing suggests Chinese leaders share the sense of urgency, if not the preferred solution.
Quiet talks between Washington and Beijing have begun to explore emergency communication channels on AI risks. The discussions, reported by the Los Angeles Times, gained momentum ahead of a planned state visit. Shared alarm over Mythos appears to have prompted both capitals to consider safeguards against unintended escalation.
Anthropic CEO Dario Amodei has met with senior White House officials including Chief of Staff Susie Wiles. The company that once warned loudly about the dangers of unchecked AI development now finds itself at the center of government strategy sessions. Its decision to withhold Mythos from China while sharing limited versions with American partners has become both a point of pride and a source of diplomatic friction.
The coming months will test whether controlled distribution can hold. New models from both American labs and their Chinese counterparts are already in development. Each iteration brings fresh capabilities and fresh temptations to loosen restrictions. Google’s detection of AI-assisted attacks shows the threat is not theoretical. The Discord breach shows containment is imperfect.
So the pressure builds. Chinese representatives will likely keep probing for openings. American firms will keep saying no to state-linked requests while they weigh how much to share with allies. And the technology itself will keep advancing, indifferent to borders or intentions. The contest that began with language models has now reached the domain of digital weapons. Neither side shows signs of slowing down.
WebProNews is an iEntry Publication