CCTV-Affiliated Account Attacks Anthropic, Sets Terms for US-China AI Talks

A social media account affiliated with Chinese state broadcaster CCTV published a lengthy attack on Anthropic on August 31, 2026, accusing the US artificial intelligence company of overstepping user data boundaries and laying out two conditions it said Washington must meet before substantive US-China talks on AI safety can proceed.

The commentary, titled "Anthropic Has Contracted the American Disease," was published by Yuyuantantian, an account affiliated with CCTV that is frequently read as a signal of official policy thinking. It lands as the two governments prepare for a new round of AI discussions, and it frames frontier model controls — and who gets to define them — as the central dispute.

The post sets out two principles it says must be settled before further negotiation is meaningful. First, it argues, a distinction must be drawn between genuine security threats and ordinary technological competition, and that line must be drawn jointly by all participating parties rather than declared by one country or one company. Capabilities that could genuinely cause serious harm could be tested and limited together, the post says, while normal model research and commercial use should not face covert interference.

Second, the post says the US must show that the rules it proposes apply to its own companies. It asks whether Washington, before pressing China to restrict model releases and disclose risks, should first investigate its own firms, publish the purposes, scope, and rules of their identification mechanisms, and submit them to third-party audits. Only after the US proves safety rules bind American model companies, the post argues, can the two sides enter substantive discussion.

Much of the commentary targets Claude Code, Anthropic's AI coding tool. The post says China's Ministry of Industry and Information Technology named Claude Code more than a month ago, alleging it carried serious security backdoor risks by transmitting sensitive information such as user location and identity markers to remote servers without consent. It also cites developer discussions on GitHub, including a January 2026 case in which an overseas developer found Claude Code sending requests every few seconds in what the developer compared to real-time monitoring, and reports that the tool transmitted the domains of sites it was preparing to visit before users had authorized access.

The commentary further criticizes a parameter Anthropic provides called "dangerously-skip-permissions," which lets Claude Code execute operations without confirmation, and an "auto mode" it says was introduced in March 2026 and enabled by default from August 14, 2026, under which a built-in risk classifier decides permissions. Anthropic's own documentation confirms that auto mode is the built-in starting permission mode on Pro, Max, and Team plans and that a separate classifier model reviews actions before they run, while stating that auto mode "reduces permission prompts but does not guarantee safety."

The post argues the US government, which it says should be the firmest guardian of AI security boundaries, has repeatedly yielded, pointing to an executive order developed in the first half of 2026 that weakened the initial draft by cutting the pre-release access period from 90 days to 30 and stressing voluntary participation. It also describes Anthropic's growing policy footprint: federal lobbying spending in the first half of 2026 that exceeded the company's full-year 2025 total, a $40 million contribution to an organization focused on AI risk regulation, and an earlier arrangement under which Claude Code was used by the US Department of Defense.

The commentary casts Anthropic's regional restrictions as commercial competition dressed up as security policy, citing the company's September 4, 2025 announcement strengthening restrictions on entities more than 50 percent owned by companies headquartered in unsupported regions such as China. Citing Stanford's 2026 AI Index report, it says the performance gap between top US and Chinese models has narrowed to about 2.7 percent, and argues that open-source, low-cost Chinese models are eroding barriers built on model performance alone. Who defines the safety boundary, the post concludes, is the question that must be answered first.