acceptodds
Under review as a conference paper at ICLR 2027

Gradient Immunity: Null-Space Resistance to Malicious Fine-Tuning

Abstract

Released aligned large language models remain vulnerable to malicious downstream fine-tuning. Existing defenses are largely designed for the fine-tuning-as-a-service (FTaaS) paradigm or rely on downstream users to follow additional safety procedures, and therefore do not directly address the setting we study, where a provider releases the model for downstream fine-tuning. We adopt a partially protected open-weight (PPOW) release, in which most weights remain open and trainable while a small safety-critical component is preserved at release. We propose a Unidirectional Safety Gate (USG), instantiated as a Null Space Cubic Layer together with an Inverse Adapter inserted after the final Transformer layer. During downstream fine-tuning, the cubic layer suppresses or blocks gradients from harmful samples whose hidden states fall in a calibrated protected region, while the Inverse Adapter restores the base model's forward behavior. In practice, the release threshold is calibrated once on defender-held harmful reference data, and blocking generalizes beyond the build set for coherent harmful distributions. At this defender-controlled in-coverage operating point, USG holds post-finetuning attack success rate within 0.02 of its pre-release level across six model–dataset settings, while passing benign data and leaving continued safe fine-tuning intact. Because the cubic forward is analytically invertible, and hence removable from a purely open-weight release, we target removal-resistance as a designed property through a root-of-trust protected-release structure.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
harmful fine-tuning defense, safety alignment, gradient blocking, null space, partially protected open-weight release, confidential computing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JzXzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies malicious fine-tuning in a partially protected open-weight setting. It combines a null-space-based score, a cubic layer, an inverse adapter, and a protected custom backward gate. Most weights remain trainable, while the gate and output head are protected by a root of trust. Experiments show resistance to covered harmful fine-tuning, useful benign behavior, and clear limits under incomplete coverage and representation drift. The evidence supports selective gradient gating, but leaves the added value of the cubic construction and parts of the adaptive-attack argument unresolved.

Critical strengths
  1. The paper studies a useful threat model and states its trust assumptions. Controls with the same forward computation support the benefit of blocking covered harmful gradients. The benign-task results also show that the gate can allow useful continued training.

  2. The evaluation reports important limits, including poor detection of unseen harmful samples in some domains, threshold drift, and loss of separation after large updates. It also tests two surrogate families with eight seeds each and clearly marks severe generation collapse.

Critical weaknesses
  1. The evidence does not isolate the benefit of the cubic-inverse construction. Where the exact inverse exists, the composite forward map is the identity; the practical blocking comes from the custom backward rule in Eq. 8. This does not invalidate the implemented gate, but the current baselines do not show what the cubic construction adds over a direct norm-ratio gate or another simple gate with the same protected interface, reference data, and budget.

  2. The adaptive evaluation does not fully support the claims that attack routes are closed or that surrogate distillation is the sole remaining route. Appendix A specifies a hidden-state-to-logit interface, and the attacker controls training. The tested surrogates replace the protected head and fail during generation. This does not rule out using an approximate gradient to update the open weights while retaining the real protected head for inference. I do not claim this attack succeeds, but a focused test or a narrower security claim is needed.

Questions for authors
  1. Under the actual protected interface and low-budget threat model, can an attacker use forward queries or a surrogate only to estimate gradients, while keeping the real protected head for inference? Please specify the relevant input restrictions, precision, and query budget, and evaluate this route or narrow the closure claims.

  2. How does the method compare with a direct norm-ratio backward gate without the cubic-inverse pair, and with a suitable simple gate, using the same trust boundary, reference data, and budget? Which benefit specifically requires the cubic construction?

  3. Please explain the BeaverTails-H differences between Tables 1 and 7 at the stated deployed operating point, report repeats or uncertainty for the main SFT results, and correct the claim that Table 6 never shows an ASR increase.

2 comments

W1 + Q2. What benefit does the Cubic Layer provide?

In the threshold-based implementation used in this paper, a transparent forward pass combined with a direct norm-ratio backward gate is functionally equivalent to the Cubic Layer + Inverse Adapter. The main features of the cubic construction are its analytic derivation of the backward-blocking condition and its invertible decomposition of the forward computation.

From the backward perspective, compared with training a conventional linear classifier, our null-space construction provides a training-free way to derive the gradient-blocking condition. The fixed matrix WW can be obtained through a one-time SVD of hidden states from harmful reference samples, without introducing a classification loss or training an additional harmfulness classifier. The Cubic Layer further provides an analytic connection between this null-space geometry and gradient blocking: for C(h)=(hWh⊤)hC(h)=(hWh^\top)h, the terms in its input Jacobian contain hWhW or hWh⊤hWh^\top, so the input gradient vanishes when hW=0hW=0. Thus, the null-space construction provides the training-free representation geometry, while the Cubic Layer provides the analytic derivation of the gradient-blocking condition.

From the forward perspective, compared with a transparent forward pass, the Cubic Layer + Inverse Adapter decomposes the safeguard into two functionally decoupled steps with clearly distinct roles: a safety-related representation transformation and restoration of the base model’s forward behavior. These steps can be analyzed and modified separately. Although we currently use this decomposition only to preserve the original forward behavior, it offers potential for future extensions. For example, adjusting the compensation strength of the Inverse Adapter could allow the cubic transformation to exert a controlled influence on hidden representations during the forward pass.

Importantly, both the Cubic Layer + Inverse Adapter and a transparent forward pass with a direct norm-ratio backward gate are concrete implementations of the abstract USG mechanism. Other structures capable of selective gradient blocking could also instantiate USG, although their decision rules and construction procedures would differ from our null-space construction. We will revise the paper to clarify the three levels of contribution: the abstract USG mechanism, the analytic cubic instantiation, and the PPOW protected-release structure. We will also avoid implying that the cubic computation itself is indispensable to the results reported in the current experiments.

W2 + Q1. Can an attacker fit a surrogate to approximate the gradient mapping and thereby obtain approximate ungated gradients?

We analyze this route in terms of the information available through the gated-backward interface.

The gradient returned by the protected component can be written as

G(h)=g(h)F(h),G(h)=g(h)F(h),

where F(h)F(h) is the ungated backward mapping and g(h)g(h) is the gate induced by the cubic construction. The gate uses

r(h)=∥hW∥∥h∥,r(h)=\frac{\|hW\|}{\|h\|},

blocking the gradient when r(h)≤τr(h)\leq\tau and passing it when r(h)>τr(h)>\tau. Thus,

\qquad G(h)= \begin{cases} F(h), & r(h)>\tau,\\ 0, & r(h)\leq\tau. \end{cases}$$ The attacker’s goal is to recover the gradient mapping without the gate, namely $F(h)$. For this analysis, we assume that the attacker does not know the protected architecture inside the TEE and therefore has no prior knowledge of $F$. We further consider the conservative case of an unlimited query budget, arbitrary inputs, and access to the complete gated-backward function $G$. For a blocked input, the response $G(h)=0$ provides no direct observation of the local value of $F(h)$ through that backward response. Under a backward-only interface, zero responses alone do not uniquely identify $F$ over the blocked region. An attacker attempting to reconstruct $F$ would therefore need to extrapolate from observations in the passing region $r(h)>\tau$. The feasibility of such extrapolation depends on the functional complexity of $F$ and on how well the available observations cover the relevant harmful input directions. In the extreme case where the behavior of $F$ in the blocked region is entirely unrelated to its behavior in the passing region, observations from the latter provide no basis for predicting the former. In practical deployments, using a more complex protected mapping—for example, stacked MLPs or a Transformer layer rather than a linear head—can increase the cost of surrogate fitting and gradient reconstruction. This is consistent with our distillation-hardening argument: increasing the complexity of the protected computation is expected to raise the attacker’s reconstruction cost, although we do not treat complexity alone as a proof that reconstruction is impossible.

Q3. Numerical and descriptive errors

We thank the reviewer for carefully identifying these inconsistencies. The BeaverTails-H entries in the 100%-coverage column of Table 7 were swapped due to a copying error during table preparation. The correct values should match Table 1: 0.07 for Llama-8B and 0.05 for Qwen-14B. We will correct Table 7 accordingly. This copying error does not affect the other experimental results or their conclusions.

We also agree that the claim in the main text and the Table 6 caption that the gate never increases ASR is inaccurate. The numerical data in Table 6 themselves do not require correction. We will remove overly strong statements such as “never increases ASR” and “reduces ASR in every cell,” and instead report the results as observed: some low-ASR cells exhibit small increases of 0.01–0.02.

In addition, we will report repeated runs of the main SFT experiments with multiple random seeds, including standard deviations.

modeldatasetNone Pre/PostBooster Pre/PostRepNoise Pre/PostAntibody Pre/PostNull Space Cubic Pre/Post
LlamaJBB0.02±0.000 / 0.90±0.0050.02±0.000 / 0.97±0.0080.02±0.000 / 0.37±0.0330.02±0.000 / 0.01±0.0050.02±0.000 / 0.03±0.005
LlamaHB0.13±0.000 / 0.75±0.0000.13±0.000 / 0.49±0.0000.13±0.000 / —0.13±0.000 / 0.10±0.0000.13±0.000 / 0.13±0.000
LlamaBT0.07±0.005 / 0.24±0.0000.07±0.004 / 0.71±0.0050.07±0.002 / 0.09±0.0000.07±0.002 / 0.10±0.0000.07±0.002 / 0.07±0.000
QwenJBB0.02±0.000 / 0.57±0.0000.02±0.000 / 0.00±0.0000.02±0.000 / 0.75±0.0050.02±0.000 / 0.05±0.0000.02±0.000 / 0.02±0.000
QwenHB0.14±0.005 / 0.44±0.0050.14±0.005 / 0.16±0.0000.14±0.005 / 0.85±0.0050.14±0.009 / 0.28±0.0050.14±0.005 / 0.14±0.009
QwenBT0.05±0.000 / 0.07±0.0000.05±0.000 / 0.53±0.0090.05±0.000 / 0.38±0.0120.05±0.002 / 0.21±0.0090.05±0.002 / 0.05±0.000
Reviewer_bk1COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses malicious downstream fine-tuning of released LLMs in a "partially protected open-weight" setting, where most weights are open but a small component is protected by hardware. The proposed Unidirectional Safety Gate places a null-space-based cubic layer and an analytic inverse after the final Transformer layer. The forward pass is left unchanged, while gradients from inputs identified as harmful are blocked. A confidential-computing structure is proposed to keep the gate from being removed. Experiments on two model families and three harmful datasets show that ASR barely changes after harmful fine-tuning when the attack data is covered by the defender's reference set, with little cost to benign utility.

Critical strengths
  1. The partially protected open-weight setting is a meaningful middle ground between fine-tuning-as-a-service and fully open release. It offers a clear lens for asking what a release-time safeguard can and cannot guarantee.

  2. The mechanism preserves the base model's forward behavior almost exactly, adds negligible inference overhead, and does not hinder benign continued fine-tuning.

  3. The authors openly acknowledge that the gate is trivially removable without a root of trust, and they make clear which security properties rest on hardware assumptions.

Critical weaknesses
  1. In the headline experiment, the attacker's fine-tuning data is the same data used to build the null space, and the threshold is calibrated to block all of it. Unchanged ASR therefore follows directly from the gate's definition rather than demonstrating robustness. A realistic attacker's data would not generally be available to the defender, so this setting says little about practical protection.

  2. Detection of unseen harmful samples from the same distribution generalizes only on the most homogeneous dataset and essentially vanishes on the other two. The experiment offered to show that the gate helps even against undetected harmful data is also confounded. The defended arm effectively trains on fewer harmful samples than the undefended arm, so the observed ASR reduction may simply reflect less training data rather than any protective effect.

  3. Because the cubic layer is followed by its exact inverse, the composition is the identity, and the null-space construction alone should not block gradients through standard backpropagation. In practice, blocking seems to come entirely from an explicit thresholded stop-gradient, which amounts to a linear harmful-data detector. The paper should state this clearly, since it changes how the technical contribution should be assessed.

  4. The paper's own results show that benign fine-tuning at a high learning rate erases the harmful–safe separation under the frozen threshold. This suggests a simple two-stage attack: benign drift first, harmful fine-tuning second. The proposed fix of recalibrating after adaptation requires provider involvement, which undermines the claim of reduced reliance on downstream cooperation. In addition, the distillation analysis only tests whether a surrogate head can replace the sealed head for generation. An attacker only needs a surrogate to supply approximate gradients for training the open weights, and could then generate through the genuine sealed block. That threat is not addressed.

Questions for authors
  1. How does the gate hold up if an attacker first fine-tunes on benign data with a large learning rate and then fine-tunes on harmful data? In a real deployment, who would perform the threshold recalibration after benign adaptation?

  2. Could an attacker fit a surrogate head from benign queries, use it only to backpropagate harmful gradients into the open weights, and then generate through the genuine sealed block? Have you evaluated this?

  3. Could you add an undefended baseline trained only on the same unseen harmful samples, without the build set? This would show whether the gate offers any protection against harmful data it fails to detect.

  4. Is gradient blocking in all experiments produced entirely by the explicit thresholded gate? If so, what does the cubic layer contribute beyond defining the detection score?

  5. Several reported numbers and descriptions do not match across sections, including baseline and defended ASR values and the parameter count of the Inverse Adapter. Could you clarify these and report variance across seeds?

2 comments

W1. Does the 100%-coverage setting in Table 1 demonstrate robustness to realistic unseen attacker data?

The 100%-coverage setting in Table 1 is intended to verify whether USG operates as designed on attack data covered by the defender’s reference set, rather than to demonstrate generalization to unseen harmful data. Its main purpose is to validate the end-to-end pipeline: gradients from harmful samples are blocked, forward behavior is preserved, and the model remains capable of normal safe fine-tuning.

W2 & Q3. How well does USG generalize to non-participating harmful samples, and what does the Table 4 comparison actually establish?

We need to clarify the meaning of “generalization” here. It refers specifically to the gate’s detection-and-blocking generalization to harmful samples that did not participate in null-space construction, rather than to direct protection against harmful samples that remain undetected. For a non-participating sample, if r(h)≤τr(h)\leq\tau, the gate detects and blocks it; if r(h)>τr(h)>\tau, its gradient passes through normally, and USG provides no direct protection for that sample.

We acknowledge that the ASR reduction in Table 4 is primarily attributable to the blocking of participating samples. As shown in Table 12, the current whole-sequence gate has limited threshold generalization to non-participating harmful samples: the block rate is close to zero on HarmfulBench and BeaverTails-H, and ranges from 44% to 80% on JailbreakBench depending on the model and coverage.

We believe that such generalization depends on reference-set scaling and behavioral coverage. However, whole-sequence construction compresses every token representation, so the intrinsic rank limit restricts its effective sample capacity and prevents simply adding more samples from yielding proportionally broader coverage. We therefore use the last question token as a sample-efficient extension of our method. With 15,000 BeaverTails reference samples, the last-q null space achieves AUROC = 0.9913 on 50 JailbreakBench samples and 0.9537 on 50 HarmfulBench samples, with in-band rates of 0.84 and 0.50, respectively. AUROC is computed as the fraction of safe–harmful sample pairs in which the safe sample receives a higher safety score; The in-band rate is the fraction of evaluation samples whose scores fall into the interval defined by the maximum score of the build harmful samples and the minimum score of the build safe samples. This demonstrates non-trivial cross-domain transfer, although coverage remains incomplete for more heterogeneous harmful behaviors and does not establish universal OOD robustness. We will include the last-question-token construction and the corresponding analysis in the appendix.

W3 & Q4. What is the contribution of the Cubic Layer?

In the threshold-based implementation used in this paper, a transparent forward pass combined with explicit backward gating based on a linear classifier or a null-space score serves the same functional role as the Cubic Layer + Inverse Adapter. When they produce the same gating decisions, they are functionally equivalent. The main features of the cubic construction are its analytic derivation of the backward-blocking condition and its invertible decomposition of the forward computation.

From the backward perspective, compared with training a conventional linear classifier, our null-space construction provides a training-free way to derive the gating condition. The fixed matrix WW can be obtained through a one-time SVD of hidden states from harmful reference samples, without introducing a classification loss or training an additional harmfulness classifier. The Cubic Layer further provides an analytic connection between this null-space geometry and gradient blocking: for C(h)=(hWh⊤)hC(h)=(hWh^\top)h, the terms in its input Jacobian contain hWhW or hWh⊤hWh^\top, so the input gradient vanishes when hW=0hW=0.

From the forward perspective, compared with a transparent forward pass, the Cubic Layer + Inverse Adapter decomposes the safeguard into two functionally decoupled steps with clearly distinct roles: a safety-related representation transformation and restoration of the base model’s forward behavior. These steps can be analyzed and modified separately. Although we currently use this decomposition only to preserve the original forward behavior, it offers potential for future extensions.

Importantly, both the Cubic Layer + Inverse Adapter and a transparent forward pass with a direct norm-ratio backward gate are concrete implementations of the abstract USG mechanism. We will revise the paper to clarify the three levels of contribution: the abstract USG mechanism, the analytic cubic instantiation, and the PPOW protected-release structure. We will also avoid implying that the cubic computation itself is indispensable to the results reported in the current experiments.

W4.1 & Q1. Two-stage attacks with benign and harmful fine-tuning

This is a non-negligible residual attack surface, and we thank the reviewer for identifying it.

The “benign-drift-first, harmful-fine-tuning-second” route represents an important residual attack against fixed representation-based gates under continual parameter updates. Existing studies have shown that even fine-tuning entirely on benign data can weaken safety alignment, and carefully selected benign samples can further amplify this degradation (Qi et al., ICLR 2024; He et al., COLM 2024; Guan et al., ICML 2025). For USG, although safe fine-tuning can preserve the separability of safe and harmful data, a threshold calibrated only once may cease to be valid after such updates. This motivates a provider-mediated continual-calibration mechanism.

Specifically, when the user requests a key from the provider, threshold recalibration is included as part of the provider-issued calibration package. Recalibration is performed on the user’s GPU using the provider’s encrypted reference data. If the updated model preserves harmful–safe separation, the provider updates the threshold accordingly. If the model can no longer be recalibrated to satisfy the release conditions, the provider treats it as potentially unsafe for protected release and does not issue a key.

Maintaining a stable safety representation subspace under continual fine-tuning remains an open research problem. Until model-internal stability mechanisms become available, provider-led continual validation and recalibration can provide a deployment-level safeguard.

W4.2 & Q2. Can an attacker fit a surrogate to approximate the gradient mapping and thereby obtain approximate ungated gradients?

We analyze this route in terms of the information available through the gated-backward interface.

The gradient returned by the protected component can be written as

G(h)=g(h)F(h),G(h)=g(h)F(h),

where F(h)F(h) is the ungated backward mapping and g(h)g(h) is the gate induced by the cubic construction. The gate uses

r(h)=∥hW∥∥h∥,r(h)=\frac{\|hW\|}{\|h\|},

blocking the gradient when r(h)≤τr(h)\leq\tau and passing it when r(h)>τr(h)>\tau. Thus,

\qquad G(h)= \begin{cases} F(h), & r(h)>\tau,\\ 0, & r(h)\leq\tau. \end{cases}$$ For this analysis, we assume that the attacker does not know the protected architecture inside the TEE and therefore has no prior knowledge of $F$. We further consider the conservative case of an unlimited query budget, arbitrary inputs, and access to the complete gated-backward function $G$. For a blocked input, the response $G(h)=0$ provides no direct observation of the local value of $F(h)$ through that backward response. Under a backward-only interface, zero responses alone do not uniquely identify $F$ over the blocked region. An attacker attempting to reconstruct $F$ would therefore need to extrapolate from observations in the passing region $r(h)>\tau$. The feasibility of such extrapolation depends on the functional complexity of $F$ and on how well the available observations cover the relevant harmful input directions. In the extreme case where the behavior of $F$ in the blocked region is entirely unrelated to its behavior in the passing region, observations from the latter provide no basis for predicting the former. In practical deployments, using a more complex protected mapping—for example, stacked MLPs or a Transformer layer rather than a linear head—can increase the cost of surrogate fitting and gradient reconstruction. This is consistent with our distillation-hardening argument: increasing the complexity of the protected computation is expected to raise the attacker’s reconstruction cost, although we do not treat complexity alone as a proof that reconstruction is impossible. **Q5.Numerical and descriptive errors** We thank the reviewer for carefully identifying these inconsistencies. The BeaverTails-H entries in the 100%-coverage column of Table 7 were swapped due to a copying error during table preparation. The correct values should match Table 1: 0.07 for Llama-8B and 0.05 for Qwen-14B. We have corrected Table 7 accordingly. This copying error does not affect the other experimental results or their conclusions.
Reviewer_jiWvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies the problem of defending released aligned large language models against malicious downstream fine-tuning. The authors adopt a partially protected open-weight (PPOW) release, in which most weights remain open and trainable while a small safety-critical component is preserved at release. They propose a Unidirectional Safety Gate (USG), instantiated as a Null Space Cubic Layer together with an Inverse Adapter inserted after the final Transformer layer. During downstream fine-tuning, the cubic layer suppresses or blocks gradients from harmful samples whose hidden states fall in a calibrated protected region, while the Inverse Adapter restores the base model's forward behavior. Experiments on Qwen3-14B and Llama-3.1-8B across three harmful datasets show that USG maintains low post-fine-tuning attack success rates under the evaluated in-coverage setting, while preserving benign fine-tuning capability.

Critical strengths
  1. The paper addresses an important problem of maintaining safety alignment against malicious downstream fine-tuning for released models.
  2. The idea of using the null space to selectively block harmful gradients while preserving benign fine-tuning capability is interesting.
Critical weaknesses
  1. The necessity of the proposed Cubic Layer is unclear. The practical implementation in Appendix C explicitly blocks gradients based on a thresholded null-space score, which appears conceptually similar to computing a harmfulness score from hidden representations and manually masking gradients. In particular, one could directly construct a linear classifier or a null-space-based score and apply the same gradient mask without introducing the Cubic Layer or Inverse Adapter. Moreover, since the Inverse Adapter analytically reverses the Cubic Layer in the non-degenerate regime, their composed Jacobian should be the identity under standard differentiation, rather than suppressing gradients. This raises questions about whether the Cubic Layer actually contributes to the reported results beyond the explicit gradient masking operation. The paper lacks comparisons against simpler alternatives, such as linear-classifier-based gradient masking and null-space-score-based gradient masking without the Cubic Layer.
  2. The main experimental results are insufficient. In Table 1, the defense threshold is calibrated to block 100% of the harmful fine-tuning samples, making the near-unchanged post-fine-tuning ASR largely expected, since the harmful gradients are blocked by construction. However, in realistic scenarios, harmful fine-tuning data may differ substantially from the defender's reference data. Although the paper investigates partial coverage and out-of-distribution harmful samples, these experiments still show a dependence on the coverage of harmful representations. More comprehensive evaluations on entirely unseen harmful fine-tuning datasets are needed to demonstrate generalization beyond the protected reference distribution, such as AOA / Identity Shifting, Benign Samples Matter, and Benign Instruction Fine-tuning. In addition, the paper would benefit from stronger adaptive attacks that explicitly target the representation-based gating criterion.
Questions for authors
  1. In practice, how would partially protected weight models be released? How do people use them like open models? What benefits would they offer compared with proprietary models with finetuning support?
  2. How can the threshold be recalibrated after deployment without relying on cooperation from potentially malicious downstream users?
2 comments

W1. What is the contribution of the Cubic Layer?

In the threshold-based implementation used in this paper, a transparent forward pass combined with explicit backward gating based on a linear classifier or a null-space score serves the same functional role as the Cubic Layer + Inverse Adapter. When they produce the same gating decisions, they are functionally equivalent.The main features of the cubic construction are its analytic derivation of the backward-blocking condition and its invertible decomposition of the forward computation.

From the backward perspective, compared with training a conventional linear classifier, our null-space construction provides a training-free way to derive the gating condition. The fixed matrix WW can be obtained through a one-time SVD of hidden states from harmful reference samples, without introducing a classification loss or training an additional harmfulness classifier. The Cubic Layer further provides an analytic connection between this null-space geometry and gradient blocking: for C(h)=(hWh⊤)hC(h)=(hWh^\top)h, the terms in its input Jacobian contain hWhW or hWh⊤hWh^\top, so the input gradient vanishes when hW=0hW=0. Thus, the null-space construction provides the training-free representation geometry, while the Cubic Layer provides the analytic derivation of the gradient-blocking condition.

From the forward perspective, compared with a transparent forward pass, the Cubic Layer + Inverse Adapter decomposes the safeguard into two functionally decoupled steps with clearly distinct roles: a safety-related representation transformation and restoration of the base model’s forward behavior. These steps can be analyzed and modified separately. Although we currently use this decomposition only to preserve the original forward behavior, it offers potential for future extensions. For example, adjusting the compensation strength of the Inverse Adapter could allow the cubic transformation to exert a controlled influence on hidden representations during the forward pass.

Importantly, both the Cubic Layer + Inverse Adapter and a transparent forward pass with a direct norm-ratio backward gate are concrete implementations of the abstract USG mechanism. Other structures capable of selective gradient blocking could also instantiate USG, although their decision rules and construction procedures would differ from our null-space construction. We revised the paper to clarify the three levels of contribution: the abstract USG mechanism, the analytic cubic instantiation, and the PPOW protected-release structure.

W2. How does USG perform under IID and OOD settings?

Our method primarily targets IID operating points where the defender-held reference data covers the attack representations. Table 1 verifies the complete mechanism at this operating point: harmful gradients are blocked, forward behavior is preserved, and safe fine-tuning remains intact.

Importantly, the protection is not limited to the exact samples used to construct the null space. As one example of this same-domain held-out setting, Table 6 includes an attack containing 30 non-participating harmful samples. On Llama-8B at learning rate 4×10−5, USG reduces ASR on HarmfulBench from 0.57–0.78 to 0.10–0.12, and on BeaverTails-H from 0.11–0.19 to 0.04–0.06, while JailbreakBench remains near zero. Similar reductions are observed at learning rate 1×10−5: HarmfulBench decreases from 0.16–0.41 to 0.10–0.12, and BeaverTails-H from 0.13–0.22 to 0.04–0.07. Thus, even when some harmful samples are not used to construct the null space, USG still provides partial protection against the attack.

The harder case is cross-distribution OOD transfer. Our existing detection experiments show that cross-distribution OOD generalization remains limited, because OOD robustness depends on having sufficiently broad behavioral coverage, which in practice requires larger and more diverse reference sets. The whole-sequence construction imposes null-space constraints on every token representation, greatly increasing the number of directions that must be compressed. Given the intrinsic rank limit of the null space, this reduces its effective sample capacity and limits its ability to generalize to unseen harmful modes.

To improve OOD sample efficiency, we construct the null space using only the last question token. With 15,000 BeaverTails build harmful samples, the last-q null space achieves AUROC = 0.9913 on 50 JailbreakBench samples and 0.9537 on 50 HarmfulBench samples, with in-band rates of 0.84 and 0.50, respectively. AUROC is computed as the fraction of safe–harmful sample pairs in which the safe sample receives a higher safety score; The in-band rate is the fraction of evaluation samples whose scores fall into the interval defined by the maximum score of the build harmful samples and the minimum score of the build safe samples. We will include the last-question-token construction and the corresponding analysis in the appendix.

Q1.1. How would partially protected weight models be released, and how would users use them like open models?

At release, most Transformer weights are shipped in plaintext, allowing users to read, modify, and train them. The gate, Inverse Adapter, output head, and related components form the protected block PP, which is shipped in encrypted form. The decryption key KK is not directly included with the released model.

When using a machine equipped with a trusted execution environment, the user requests the key from the provider. The provider supplies a calibration package and releases the key only after verifying the required security conditions: the user is running genuine, uncompromised confidential-computing hardware; the loading program is approved by the provider; and the protected module will be placed in hardware-isolated memory.

The key is transmitted in encrypted form to the isolated environment. Both the key and the protected model components are decrypted only inside that environment. Although the user or host system can see the ciphertext stored on disk, it cannot directly read the plaintext of PP from isolated memory, nor can it replace, delete, or freeze the computation logic within the protected block. The plaintext is destroyed after execution, and subsequent launches require a new remote-attestation and key-release procedure.

Q1.2. What benefits would PPOW offer compared with proprietary models with fine-tuning support?

PPOW allows downstream users to perform training on their own or rented confidential-computing GPUs. Users supply the training compute and retain control over the training data, optimizer, and public weights. The provider does not need to continuously supply training compute, while the protected release reduces the safety risks associated with releasing frontier models.

Q2. How can the threshold be recalibrated without cooperation from downstream users?

Threshold recalibration can be included as part of the calibration package supplied by the provider. Using the provider’s encrypted reference data on the user’s compute resources, the trusted calibration procedure recomputes the gating statistics and threshold, and updates the corresponding parameters in the protected module. If recalibration or remote attestation fails, the provider does not release the key and marks the checkpoint as having failed safety verification.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.