Anthropic Rejects China’s Claude Code Backdoor Claim
Anthropic denies China’s backdoor warning on Claude Code, while new research tackles MoE inefficiencies and conformal prediction gaps.
The story of the day
Anthropic pushed back against China’s warning of a security backdoor in Claude Code, reiterating its policy banning access to its models and coding tools in the country. Beijing’s alert described a "hidden backdoor" in the tool, but Anthropic framed it as a misinterpretation of standard functionality. The exchange underscores the growing friction between U.S. AI providers and Chinese regulators, with builders caught in the middle: those relying on Claude Code must now weigh compliance risks against tooling benefits, while China’s developers face another barrier to Western AI.
New tools & launches
TriRoute — Open-source framework for unified adaptive attention, expert selection, and KV-cache allocation in LLMs, reducing inference costs without sacrificing quality. [arXiv:2607.06601]
Models & research
TriRoute (arXiv:2607.06601) — Proposes a single controller to jointly optimize attention resolution, expert routing, and KV-cache precision, addressing the inefficiency of isolated conditional computation methods. Open-weight, it claims near-linear scaling for MoE models by coupling these decisions per-token, per-layer.
Marginal Conformal Prediction Flaw (arXiv:2607.06605) — Reveals standard conformal prediction fails on imbalanced datasets, with minority class coverage dropping to 4.2% in toxicity prediction despite meeting global 90% targets. The fix: class-conditional calibration, tested across random forests, GNNs, and frozen chemical LMs.
Industry pulse
China’s warning about Claude Code’s alleged backdoor highlights the escalating trust gap between Western AI tools and non-Western regulators. Anthropic’s response—reiterating its access ban—avoids direct technical rebuttal, leaving developers to parse the risk themselves. The episode mirrors last month’s Alibaba ban on Claude Code over similar concerns, suggesting a pattern: regulatory scrutiny now targets tooling as much as models.
Stat of the day
Conformal prediction’s minority-class coverage can plummet to 4.2% in imbalanced datasets (e.g., clinical toxicity), despite hitting 90% global coverage targets. [arXiv:2607.06605]
Get the full analysis — subscribe to Gradient Pro.
