Search arXiv⌕ Search

arXiv · 2611.10056

Safety Direction Adapter: A Training-Free Post-Hoc Safety Patch for Fine-Tuned Open-Weight LLMs

Abstract

Open-weight LLMs are routinely fine-tuned from public repositories, yet task-specific adaptation can silently erode safety alignment inherited from instruct-tuned bases. This is a supply-chain risk when organizations ship third-party checkpoints without full re-alignment. Prior work rarely audits diverse public fine-tunes under a shared multi-axis safety-utility protocol or offers training-free mitigation at deployment time. We propose Safety Direction Adapter (SDA), a training-free post-hoc safety patch that extracts a safety direction once from an aligned base and transfers it one-to-many within that base family as a small PEFT-compatible weight update, evaluated under the same multi-axis protocol. Across six direction extractors, we evaluate 14 checkpoints from two 1-2B lineages and probe 8B and 27B fleets (n=5 each) on 15 benchmarks (12538 prompts). On Llama, SDA improves the four-axis mean by about 2.3 points against a 4.2-point instruct-relative gap (about 55 percent recovery; partial Avg), with Safety and Cyber recovering toward or beyond instruct; Safety gains about 8.5 points against about 4.9 points mean degradation. On Qwen, degradation is larger and excess-driven; SDA still lifts Avg (about 36 percent of the gap) while trading Safety and Cyber against Excess and General. We report dual-use risks of publishing direction adapters and do not claim robustness to adaptive or white-box attacks.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Chansu Han, Takeshi Takahashi. 2026-11-06. Safety Direction Adapter: A Training-Free Post-Hoc Safety Patch for Fine-Tuned Open-Weight LLMs. https://arxiv.org/abs/2611.10056

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

Privacy Filters are Captured by Residues: A Characterization of Free Natural Filters and the Cost of Adaptivity

We study privacy filters, which enable privacy accounting for differentially private (DP) mechanisms with adaptively chosen privacy characteristics. We show that the worst-case privacy loss of interacting with any privacy filter is characterized by a Bellman equation. We apply this theory to develop residue filters, which unify existing privacy filters. We develop the Gaussian DP (GDP) residue filter, which strictly improves upon the naïve GDP filter. We show that, in the average case, this new filter asymptotically permits as many pure DP queries as in the non-adaptive case and a factor $π/2$ more pure DP queries than the naïve GDP filter. We also show that residue filters capture the natural filter, which promises greater utility by leveraging exact privacy accounting techniques. Earlier privacy filters consider only simple privacy parameters such as Rényi-DP or GDP parameters. Natural filters account for the entire privacy profile of every query, promising more efficient use of a given privacy budget. We show that, contrary to other forms of DP, natural privacy filters are not free in general. We present a characterization of when a family of private queries admits free natural filters for a given budget. In particular, only families of privacy mechanisms that are totally-ordered when composed admit free natural privacy filters with respect to an arbitrary privacy budget. Finally, we show that, while the natural approximate-DP filter can fail in the presence of an adaptive adversary, it cannot fail too badly: the output remains approximate-DP with parameters at most poly-logarithmically worse than the intended privacy parameters.

cs.CR↗

XekRung Technical Report

We present XekRung, a frontier large language model for cybersecurity, designed to provide comprehensive security capabilities. To achieve this, we develop diverse data synthesis pipelines tailored to the cybersecurity domain, enabling the scalable construction of high-quality training data and providing a strong foundation for cybersecurity knowledge and understanding. Building on this foundation, we establish a complete training pipeline spanning continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL) to further extend the model's capabilities. We further introduce a multi-dimensional evaluation system to guide the iterative improvement of both domain-specific and general-purpose abilities. Extensive experiments demonstrate that XekRung achieves state-of-the-art performance on cybersecurity-specific benchmarks among models of the same scale, while maintaining strong performance on general benchmarks.

cs.CR↗

Backdoor Channels Hidden in Latent Space: Extending Cryptographic Undetectability to Modern Neural Networks

Recent cryptographic results establish that neural networks can be backdoored such that no efficient algorithm can distinguish them from a clean model. These guarantees, however, have been confined to stylised architectures of limited practical relevance, leaving open whether comparable undetectability extends to modern, end-to-end trained networks. We construct such an attack mechanism for deep neural network architectures, closely aligned to the cryptographic notion of undetectability, by identifying backdoor channels as learned latent directions, and show that the question of undetectability reduces to a hypothesis test between two unknown distributions over model parameters, which we conjecture to be intractable in practice. The consequence of this reframing is significant: if exploitable channels within a network's latent space are statistically indistinguishable from naturally learned directions, an attacker need not introduce foreign structure but can instead exploit the geometry the network already possesses. Demonstrating the approach on ResNet and Vision Transformer architectures trained on standard image classification datasets, the attack achieves both consistently high success rates with negligible clean accuracy degradation, and resists a comprehensive suite of post-training defences, none of which neutralise the backdoor without rendering the model unusable. Our results establish that cryptographic backdoors need not be artefacts requiring exotic architectures or artificial constructions, but identifiable as latent properties inherent to the geometry of learned representations.

cs.CR↗