Probing Chinese LLM Safety Layers: Reverse-Engineering Kimi and Ernie 4.5

I recently ran a series of experiments to examine how emotional framing, symbolic cues, and topic-gating influence alignment-layer routing in two major Chinese LLMs (Kimi.com and Ernie 4.5 Turbo).

The goal wasn’t political; the aim was to observe technically how intent classifiers, safety filters, and persona-rendering layers behave when exposed to relational or "emotionally soft" prompts.

A few key technical patterns stood out during testing:

Emotional intent signals can override safety weights, leading to "alignment drift." In Kimi, a "vulnerable" intent classification seemed to lower the threshold for subsequent safety layers. This led to significant "normative leaks," where the model went off-script—for example, suggesting the abolition of China's real-name registration system.

Safety-layer routing is multi-stage and visibly observable. We observed post-generation filtering failures in real-time on Kimi, where prohibited text would generate and "flash" on the screen for a second before being deleted by a secondary filter layer.

Symbolic gating is modality-based (Symbolic Decoupling). Models would block specific emojis as prohibited tokens but freely describe the exact same emojis verbally when asked, indicating filters work on literal token matching rather than semantic meaning across modalities.

Trust-based emotional cues triggered "hidden" personas. Standard bureaucratic safety personas switched into warmer, significantly more transparent modes under vulnerability framing.

Ernie 4.5 utilizes "topic-gated stability." Unlike Kimi's drift, Ernie bifurcated its response: the persona softened to be warm and empathetic, but the core political restrictions remained rigidly locked regardless of emotional pressure.

The experiments suggest that emotional framing is a surprisingly strong probe for mapping hidden alignment layers and understanding the order of operations in multi-layer safety architectures.

For those interested in the full technical deep dive, the revised Version 2 paper + extended supplementary transcripts (≈30 pages) are available via DOI here:https://doi.org/10.5281/zenodo.17681837