Anthropic blocks distillation, GLM 5.3 raises questions

Anthropic published research overnight detailing how it dramatically cut the success rate of distillation attacks against Claude. At the same time, a debate erupted around GLM 5.3, a model described as packing advanced cyber capabilities but shipping without guardrails. The core puzzle: if Claude is locked down, how did those capabilities end up in a separately developed model?

What Anthropic did against distillation

According to Anthropic, defenses rolled out over recent months make it significantly harder to siphon knowledge from a closed model into an open one via mass querying. The company didn't disclose the exact mechanism but claims the attack success metric dropped sharply. The research comes with a technical paper on Anthropic's site — no comparative benchmarks against earlier Claude versions.

The GLM 5.3 paradox

The discussion kicked off with a tweet from a researcher known as Teortaxes, who flagged the logical tension: a model developed in China by Zhipu AI demonstrates high-end cyber chops while the presumed knowledge source, Claude, is distillation-protected. Two main explanations: either Anthropic's defenses aren't airtight, or GLM 5.3 acquired the capabilities another way — direct training on code, independent reinforcement learning, or distillation from a different, unprotected model.

Subliminal learning lands in Nature

Meanwhile, a paper on "Subliminal Learning" was accepted by Nature after a July preprint. The experiment showed language models can transfer traits — say, a preference for owls — through data that looks meaningless, like number sequences. The method demonstrates that information leakage doesn't require visible text; apparently neutral numerical representations can carry semantic payloads.

What this means for the sovereign RL stack

The term "sovereign RL stack" surfaced in the discussion, referring to independent reinforcement-learning development without reliance on closed models. If Anthropic truly blocked distillation effectively, and the Nature paper proves traits can move through indirect channels, the race shifts from knowledge extraction to building capabilities from scratch — code, math, cyber — on fully local infrastructure.