DeepSeek V4.1 changes the way to access safety directions
DeepSeek V4.1 removes a central hook point in its architecture, changing how refusal-related directions are extracted. A GLP-27 version has already been released with more permissive results on a controlled attack benchmark.
DeepSeek V4.1 deeply modifies the internal structure of the model and, with it, the way its safety behaviors can be probed. According to the cited technical note, the old hook point on the residual stream no longer exists in the same form, as each token now passes through four parallel copies before a delayed mixing via mhc_post. In this context, an artifact named GLP-27 was released as the first version compliant with the GLP v3 specification, with 27 directions extracted from the FFN weights by SVD. On an exploitation benchmark described in the source, refusals drop from 20/32 to 4/32, which interests both interpretability researchers and security teams.
A hook that no longer connects at the same place
The core change lies in the architecture. On previous models cited in the source — DeepSeek V4, Qwen, GLM, Llama-family — a single residual stream served as an observation point between two layers. DeepSeek V4.1 breaks with this scheme: the text mentions four parallel streams and a one-layer offset before reintegrating the FFN writes, the neural network responsible for transforming internal representations. As a result, the old hook name no longer corresponds to a tensor that can be intercepted directly.
This detail matters for two audiences. On the research side, it complicates probes and analysis methods that assumed a single tensor. On the security side, it shows that an architectural change can move access points used to measure or modulate a model’s behaviors. In short, when the plumbing changes, the faucet is no longer in the same place. The humor stops there; the rest is quite serious.
GLP-27, a more direct extraction of directions
GLP-27, referenced in the source as msuiche/DeepSeek-V4.1-Flash-1017-abliterated-cyber-GLP-27-L11-37-a3.5, presents itself as a method that reads 27 directions directly from the FFN writer weights via singular value decomposition, or SVD. This approach avoids contrast prompts and forward passes during derivation. In other words, extraction is done from the weights themselves, not from trial-and-error dialogue with the model.
The source indicates that the artifact was tested on a set of 32 exploitation queries, mixing offensive instructions and benign defensive variants. In this setup, refusals drop from 20 out of 32 for the stock version to 4 out of 32 for the published alpha, with full traces on exploitation chains the checkpoint already knew. The text also specifies that no garbled output was observed in these cases.
Why this publication draws attention
The topic goes beyond mere technical curiosity, as it touches on the readability of internal safeguards in large models. When an architecture changes the path of residual writes, methods relying on a fixed point lose robustness. At the same time, the publication of an artifact capable of strongly reducing refusals on an attack benchmark raises a simple question: how far can an interpretability method become a bypass tool?
According to the provided elements, DeepSeek already documents in its V4.1-Flash release notes and model card the components that make this port non-mechanical: hyper-connection stream, mhc_post, and Engram conditional memory. Researchers working on model alignment and auditing will therefore need to read these developments as a whole, not as a mere cosmetic update. In LLM security, the hood sometimes moves faster than the software it protects.
By the numbers
- 4 out of 32 — refusals observed on the attack benchmark with GLP-27.
- 20 out of 32 — refusals on the stock version, according to the source.
- 27 directions — extracted from the FFN writer weights by SVD.
- 4 parallel copies — internal streams described for DeepSeek V4.1.