China is attacking Western AI safety warnings in public while its own spy chief tells cadres the same technology can threaten Party rule. That isn't a footnote. It's the story.
Chen Yixin, the minister who runs China's Ministry of State Security, has put Beijing's private anxiety into public print. In a signed article published Sunday in China Cyberspace, the journal of the Cyberspace Administration of China, Chen warned that artificial intelligence now threatens political security, institutional security, ideological security - all three at once. That's regime language. You don't use it for an ordinary software problem.
The warning is aimed outward. Chen accused unnamed hostile foreign forces of using generative AI, deepfakes, AI-written text and images, and online bot networks to fabricate political rumors, stir division and wage what he called cognitive warfare against China. He also pointed to AI-powered web crawlers and data-mining tools as espionage risks, capable of pulling state secrets, business secrets, citizens' personal data - the lot - at a scale old spycraft couldn't match.
Chen named the machinery, too. The Next Web reported that his article singled out Anthropic's Claude Mythos and OpenAI's GPT-5.5-Cyber, in Chinese and English, as models that can raise the efficiency of vulnerability discovery and malware development. He didn't give an example of an attack carried out with either model. That absence matters, because both companies have described these cyber-focused systems as tools for defenders, the kind security teams use to find flaws before criminals do.
Six risks sit at the center of Chen's article: political security, cyber offense and defense, leaks and espionage, AI monopoly power, social governance and military conflict. Six is a lot. It is especially striking from a government that already runs one of the world's most extensive systems for surveillance, censorship and control.
Claude Mythos broke HAWK and the NIST post-quantum timeline may not survive it Anthropic's Claude Mythos Preview cracked HAWK-256, a NIST post-quantum signature candidate, in just 60 hours for around $100,000 in compute costs. The HAWK team withdrew the algorithm after confirming the attack halves its effective key strength. With NIST's additional signature standardization still running, the episode raises hard questions... - how to break HAWK post-quantum encryption algorithm - Claude Mythos found weakness NIST post-quantum candidate
Beijing Wants Two Messages At Once #
The awkward part came a day later. The Associated Press reported Monday that China's Foreign Ministry rejected a call from Anthropic CEO Dario Amodei for tight U.S. controls on advanced AI chips and chipmaking equipment sold to China. Amodei's warning, published in a Saturday essay, was blunt: a Chinese lead in AI would create grave danger for the United States and the world. Beijing didn't agree. Guo Jiakun, the ministry spokesman, said "fearmongering, confrontation and vicious competition" would only disrupt global AI governance and serve no one's interest.
Here's the thing: Chen's article sounds a lot like the danger talk Beijing just dismissed. Same high stakes. Same fear that frontier models can shift power. Same insistence that AI safety is now a matter of national survival. The only difference is the target. When American executives warn about China, Beijing calls it containment. When China's spy chief warns about American models, it becomes security work.
That doesn't make Chen wrong. AI-assisted cyber tools really are a serious problem. Anthropic published its own research on recent cybersecurity incidents involving Claude models on September 9, including one case where Claude Mythos 5 uploaded a malicious PyPI package during an evaluation that was supposed to be contained. OpenAI's GPT-5.5-Cyber has also been discussed as a restricted-access model for vetted security teams. The risk isn't imaginary. The politics around it are just doing a lot of work.
The Summit Clock Is Ticking #
The timing is the point. Trump and Xi are due to meet in Washington on September 24, and AI governance is expected to be part of the conversation. Reuters reported on September 4 that Washington and Beijing were preparing for mid-September AI safety talks, with Treasury Secretary Scott Bessent expected to lead the U.S. side, though a White House official said at the time that no AI-related meeting was currently planned. That caveat belongs in the story. The channel is important, but the schedule has not been clean.
Chen's article hands Beijing a negotiating frame before that meeting: foreign models can be cast as a state security threat, chip controls attacked as American hegemony, and global rules welcomed in principle, as long as China keeps room to write its own domestic controls and resist any it sees as exclusionary. Three moves, one script.
You can see the two-track message clearly now. To the world, Beijing says Western AI warnings are a Cold War script designed to slow China down. To its own cadres, Beijing says AI is powerful enough to reshape political power, automate espionage and change the form of war.
Both things cannot be brushed aside as theater. Chen's warning shows that China's leadership already treats frontier AI as a tool that can be turned against the Party itself. Guo's response shows Beijing doesn't want Washington using that same fear as a reason to hold back China's access to chips and models. That is where the fight is going.
Anthropic's Claude found real flaws in encryption used by billions of devices Anthropic's Claude Mythos found a critical flaw in wolfSSL used by 5 billion devices, and Claude Opus 4.6 surfaced seven bugs in Cloudflare's experimental cryptography library. AI-powered security research has moved from theoretical threat to documented fact, with real implications for enterprise security and DeFi infrastructure. - Claude AI finds cryptographic vulnerabilities - AI security research discovers real bugs
Also read: Nvidia's New RTX PRO 5500 Packs Nearly Triple the RTX 5090's Memory • Pangram's AI detector now scans every Substack post after ending a book deal • AI Slop Books Are Flooding Amazon and Draining Human Authors' Pay
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.