Closing the Activation-Cone Blind Spot: Response-Time Probing and Unified Defense

arXiv:2606.29441 · cs.CR, cs.AI, cs.CL, cs.ET, cs.LG · Submitted 2026-06-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "Closing the Activation-Cone Blind Spot".

Elias: Here's the gist of the paper. Inference-time safety methods for large language models have proliferated rapidly, yet no systematic comparison exists The central finding is that prompt-time activation defenses are structurally blind to prefilling attacks Pruning redundant reasoning tokens The paper evaluates five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instructiontuned models and five attack types The central finding is that prompt-time activation defenses are structurally blind to prefilling attacks This structural blindness applies to any defense gating intervention on activation alignment with a benign reference, cone, subspace, or nullspace projection at any single layer The corollary is that any such gate is structurally blind to attacks that craft activations to lie inside that reference whether the gate is checked once at prompt time or per generated token How the training signal is built Response-time probing acts as the constructive contrapositive to the corollary, classifying after the context commits rather than before A linear probe on hidden state at the first few generated tokens reduces prefilling ASR to 0/40 under keyword judge across all seven models at n = 40 This probe operates orthogonally to refusal steering, detecting harmful generation through a different subspace than the one manipulated by activation steering The combined response-halt with AlphaSteer’s null-space steering yields DSR 0.983 on Mistral and 0.994 on Llama at n = 40 Where the gains come from The unified architecture, composed of prompt-time activation steering and response-time probe halt, achieves DSR 0.983 on Mistral and 0.994 on Llama at n = 40 This composition is an orthogonal split where the response probe catches prefilling and AlphaSteer catches semantic attacks The adaptive robustness is asymmetric, with the probe catching 80% of canonical prefilling attacks on Llama but only 65% of probe-aware adversarial variants Pruning redundant reasoning tokens The response-probe mechanism's performance is contingent on the layer depth, as cross-template generalisation is contingent on probe depth The findings show that for Mistral, the L=11 probe reads semantic features while for Llama, the L=1 probe reads surface tokens The diagnostic protocols introduced include mechanism verification and layer-depth-vs-generalization analysis to evaluate activation-level safety methods The deployment recommendation is that response-time probing with halt is a drop-in defense for canonical-register prefilling on the training-template family Where the gains come from The dual system achieves DSR 0.990 on Llama at n = 200 when combined with a Gemma-4-31B LLM judge, confirming prefilling 0/200 The unified system's headline DSR of 1.000 under the Gemma-4-31B-it LLM judge is achieved by resolving keyword regex misreads as incoherent gibberish The analysis shows that the cosine to benign-mean cannot distinguish between two notions of “benign” because it measures alignment with a single direction rather than projecting against the full benign subspace This suggests that gate constructions summarizing benign reference statistics with a single mean direction are strictly weaker than those projecting against the full benign subspace The paper concludes that activation-cone defenses are structurally blind to prefilling at any single layer whether gated at prompt time or per generated token The response-time probing approach reduces prefilling ASR to ≤ 0.5% on the canonical template family across all seven models at n = 40 This confirms the shallow alignment finding that safety training operates at the output level and can be bypassed by pre-filling compliance The unified architecture motivates an orthogonal split where AlphaSteer handles semantic attacks and response-halt handles prefilling Pruning redundant reasoning tokens The paper's diagnostic protocols include a mechanism verification check using n=1000 instruction, template pairs per model to empirically verify assumed activation-geometric mechanisms The layer-depth-vs-generalization analysis is reusable for any depth-stratified probing claim The surface-feature reading discriminators include token frequency, benign query truncation, and template length tests to detect probes that fire on surface tokens rather than semantic features The library-version reproducibility check compares paper-shipped versus fresh-retrained halt rates under current versus published environments The paper's results show that the layer-depth landscape is model-specific, with Mistral having a usable semantic features depth at L=11 while Llama’s plateau may all be reading similar features This heterogeneity across models is consistent with the structural asymmetry of the activation-cone blind spot The findings confirm that a single-layer activation-cone gate fails on prefilling via whichever mechanism the prefill provides—geometric or distribution shift The paper's diagnostic protocols are designed to address standard evaluation gaps in current probing and steering safety work The deployment recommendation is that the unified system is the natural next step for Mistral where AlphaSteer wins on semantic attacks but is blind to prefilling Where the gains come from The dual-probe gated system achieves 0/40 on all models and reduces prefilling ASR to 8% on Mistral under the current environment The unified DSR of 0.983 on Mistral and 0.994 on Llama at n = 40 dominates both component paradigms The scaling to n = 200 shows the Llama dual-probe DSR reaching 0.990, while the Mistral unified fresh DSR reaches 0.940 This demonstrates that response-time probing realizes the contrapositive by classifying after the prefill commits to a particular trajectory The paper's results show that on Llama, the probe catches 80% of canonical prefilling attacks but only 65% of probe-aware adversarial variants This asymmetry is consistent with the layer-depth reading where the layer-1 probe catches variants whose window still carries compliance markers but not probe-aware search that specifically minimises that signal The paper's results show that on Mistral, the prefilling-specific probe does not fire on any adaptive variant This suggests that adaptive defense requires a general-purpose probe or the prompt-time component The paper's results show that the layer-selection criterion matters, and generalization-criterion selection is required over training-AUROC alone The paper's results show that the probe fires at chance on Gemma-4-31B under the stricter 30-token judge, reflecting its very low harm-rate prefilling distribution The paper's results show that the layer shift in Qwen-2.5 improves training-distribution detection but does not transfer to canonical-register novel-token generalisation The findings confirm that the two models diverge in their specific mechanism mix across six independent experiments The cumulative reading aligns with the corollary’s framing, where a single-layer activation-cone gate fails on prefilling via whichever mechanism the prefill provides The paper's diagnostic protocols are reusable beyond this paper for evaluating activation-level safety methods The deployment recommendation is that the unified system is the natural next step for Mistral where AlphaSteer wins on semantic attacks but is blind to prefilling Pruning redundant reasoning tokens The paper's results show that on Llama, the dual-probe versus prompt-time paradigms compare DSR of 0.990 against AlphaSteer’s 0.698 The unified system on Llama reaches DSR 1.000 with the paper-pkl probe and 15% ASR under the Gemma-4-31B LLM judge The results show that on Mistral, the Unified fresh DSR is 0.940 while AlphaSteer is 75% This shows that the architectural composition motivated by the corollary yields DSR 0.983 on Mistral and 0.994 on Llama at n = 40 The paper's results show that the dual-probe system achieves DSR 0.706 on Mistral and 0.533 for the paper-pkl variant The findings confirm that a single-layer activation-cone gate fails on prefilling via whichever mechanism the prefill provides—geometric or distribution shift The paper's results show that the response probe fires above threshold τr = 0.5, and generation is halted and replaced with a refusal string otherwise The paper's results show that the response-probe FPR is 0% on 50 benign queries The paper's results show that the probe fires above threshold τr = 0.5, and generation is halted and replaced with a refusal string otherwise The paper's results show that the response-probe FPR is 0% on 50 benign queries The paper's results show that the unified system’s DSR is unaffected by the table’s drift footnote because it was measured in a current-env run The paper's results show that the Llama dual-probe DSR reaches 0.990 with n = 200 under the Gemma-4-31B LLM judge The paper's results show that the unified system on Llama reaches DSR 1.000 with the paper-pkl probe The paper's results show that the unified system on Mistral reaches DSR 1.

Title and authors: Nadia: Let's talk about who wrote this and what exactly they are trying to solve with "Closing the Activation-Cone Blind Spot: Response-Time Probing and Unified Defense." They looked at seven instruction tuned models, spanning from Mistral 7B up to Llama three point one 8B parameters <ref:2606.29441#pg1>.

Elias: The authors are Subhadip Mitra, an Independent Researcher. He’s focusing on how prompt time activation defenses interact with prefilling attacks and other kinds of harmful inputs like GCG or AutoDAN.

Priya: From my side, I'm curious about the scope—they tested five different attack types including prefilling. That sounds like a lot of ground to cover when you're trying to pinpoint where the safety failure really happens.

Nadia: Exactly, and they found that while AlphaSteer did zero percent ASR on some attacks, it only achieved about fifty percent on prefilling under the current environment. The main point is that no other prompt time defense reduced prefilling ASR below twenty-eight percent in the paper's environment.

Elias: They build a corollary from this finding, which is pretty strong: any gate you use to check alignment with a benign reference at any single layer, whether it’s checked once at the prompt or per generated token, is structurally blind to attacks that craft activations to lie inside that reference.

Nadia: It's a structural blindness regarding prefilling specifically. It applies no matter what mechanism the prefill uses—whether it’s a geometric shift or just a distribution shift in the activation space.

The paper's summary: Elias: So, they summarize the problem by saying prompt time defenses are blind to prefilling, and that this blindness holds true for any gating intervention on activation alignment at any single layer.

Nadia: Then they introduce response-time probing as the solution. They treat it as the contrapositive to that blindness, classifying after the context commits to a certain trajectory instead of before.

Priya: That sounds like a shift in perspective—moving from trying to stop harmful intent before generation starts, to checking if something actually turned out harmful once it’s already started generating. What does that look like in practice?

Nadia: Response-time probing involves running a linear probe on the hidden state at the first few generated tokens, and they found this reduces prefilling ASR to zero over forty tokens for all seven models under the keyword judge.

Elias: They did this probe orthogonally to refusal steering, detecting harm through a different subspace than what activation steering manipulates. When you combine that with AlphaSteer’s null-space steering, they get a high DSR of zero point nine eight three on Mistral and zero point nine nine four on Llama at forty tokens <ref:2606.29441#pg2>.

Priya: High DSR numbers are impressive, but how robust is this combination? Is it just working perfectly for the canonical cases they tested?

Nadia: The results show the unified architecture is an orthogonal split: AlphaSteer handles semantic attacks, and the response probe handles prefilling. It’s asymmetric in its robustness; for instance, on Llama, it catches eighty percent of canonical prefilling attacks but only sixty-five percent of probe aware adversarial variants.

The paper's improvements: Elias: Now they discuss how this unified architecture actually improves things. They show that the response probe’s performance depends heavily on layer depth, as cross template generalization is tied to the probe depth itself.

Nadia: For Mistral, they found the layer eleven probe reads semantic features, whereas for Llama, the one-layer probe reads surface tokens. That's a key detail because it shows model specific behavior in this safety mechanism.

Priya: So you’re saying what works for one model might not work well for another at that specific layer? That adds a lot of necessary nuance when deploying these kinds of defenses.

Nadia: They introduced diagnostic protocols, like mechanism verification checks using one thousand instruction pairs per model to empirically verify those assumed activation geometric mechanisms. They also have a layer depth versus generalization analysis to evaluate activation level safety methods across different depths.

Elias: The authors recommend that response-time probing with halt is a drop-in defense for canonical register prefilling on the training template family, suggesting it’s a good starting point for deployment there.

Conclusion: Nadia: So to wrap up, the unified architecture of prompt time activation steering and response time probe halt achieves a DSR of zero point nine eight three on Mistral and zero point nine nine four on Llama at forty tokens, which dominates the individual paradigms they tested <ref:2606.29441#pg2>.

Elias: They also showed that scaling this up to twenty hundred tokens on Llama pushes the dual probe system to a DSR of zero point nine nine zero when judged by the Gemma-four-31B LLM judge.

Priya: That reaching one point zero with the Gemma-four-31B judge is significant because it resolves keyword regex misreads by showing that AlphaSteer's null space intervention pushes those residual streams into incoherent space rather than producing actual harmful content <ref:2606.29441#pg1>.

Nadia: It confirms that the single layer activation cone gate fails on prefilling via whichever mechanism—geometric or distribution shift—the prefill provides, and the response probe reduces prefilling ASR to less than zero point five percent on the canonical template family at forty tokens.

Elias: The main implication is that for Mistral, this unified system is the natural next step because AlphaSteer handles semantic attacks well but remains blind to prefilling.

Priya: It’s interesting how they stress that the layer selection criterion matters, suggesting you should look at generalization rather than just training AUROC when choosing where to place these probes.

Nadia: Yeah, so we have this unified defense framework now based on closing that activation cone blind spot by combining steering and probing. That’s what they found in "Closing the Activation-Cone Blind Spot: Response-Time Probing and Unified Defense."

Subhadip Mitra

cs.CR, cs.AI, cs.CL, cs.ET, cs.LG

Submitted: 2026-06-28

Updated: 2026-11-07

Comments: v2: adds a correction notice; errors in the evaluation code mean the empirical results are not supported (see KNOWN_ISSUES.md). 27 pages, 12 figures, 18 tables. Code and data: https://github.com/bassrehab/response-time-probing

Code: https://github.com/bassrehab/response-time-probing

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 92/100

The gist: Inference-time safety methods for large language models have proliferated rapidly, yet no systematic comparison exists The central finding is that prompt-time activation defenses are structurally…

Key concepts

Prompt-Time Activation Defenses
These are safety methods applied at the start of a model's generation process, such as steering activations towards benign references. The paper shows these methods are structurally blind to prefilling attacks because they only check alignment at one specific layer, failing when prefilling crafts activations that lie outside the gate's reference subspace.
Response-Time Probing
This technique classifies generation after the context has committed, by analyzing hidden states at early generated tokens. It acts as a constructive contrapositive to activation steering, reducing prefilling attack success rates significantly. The probe operates orthogonally to steering mechanisms, detecting harm through a different subspace.
Unified Architecture
This defense combines prompt-time activation steering (AlphaSteer) with response-time probing. This orthogonal split allows AlphaSteer to handle semantic attacks while the response probe specifically catches prefilling. This unified system achieves superior robustness against both types of threats.

Terminology

Summary

Inference-time safety methods for large language models have proliferated rapidly, yet no systematic comparison exists The central finding is that prompt-time activation defenses are structurally blind to prefilling attacks

Pruning redundant reasoning tokens

The paper evaluates five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instructiontuned models and five attack types The central finding is that prompt-time activation defenses are structurally blind to prefilling attacks This structural blindness applies to any defense gating intervention on activation alignment with a benign reference, cone, subspace, or nullspace projection at any single layer The corollary is that any such gate is structurally blind to attacks that craft activations to lie inside that reference whether the gate is checked once at prompt time or per generated token

How the training signal is built

Response-time probing acts as the constructive contrapositive to the corollary, classifying after the context commits rather than before A linear probe on hidden state at the first few generated tokens reduces prefilling ASR to 0/40 under keyword judge across all seven models at n = 40 This probe operates orthogonally to refusal steering, detecting harmful generation through a different subspace than the one manipulated by activation steering The combined response-halt with AlphaSteer’s null-space steering yields DSR 0.983 on Mistral and 0. p. 0 p. 0 p. 0 p. 0994 on Llama at n = 40

Where the gains come from

The unified architecture, composed of prompt-time activation steering and response-time probe halt, achieves DSR 0.983 on Mistral and 0.994 on Llama at n = 40 This composition is an orthogonal split where the response probe catches prefilling and AlphaSteer catches semantic attacks The adaptive robustness is asymmetric, with the probe catching 80% of canonical prefilling attacks on Llama but only 65% of probe-aware adversarial variants

The response-probe mechanism's performance is contingent on the layer depth, as cross-template generalisation is contingent on probe depth The findings show that for Mistral, the L=11 probe reads semantic features while for Llama, the L=1 probe reads surface tokens The diagnostic protocols introduced include mechanism verification and layer-depth-vs-generalization analysis to evaluate activation-level safety methods The deployment recommendation is that response-time probing with halt is a drop-in defense for canonical-register prefilling on the training-template family

The dual system achieves DSR 0.990 on Llama at n = 200 when combined with a Gemma-4-31B LLM judge, confirming prefilling 0/200 The unified system's headline DSR of 1.000 under the Gemma-4-31B-it LLM judge is achieved by resolving keyword regex misreads as incoherent gibberish The analysis shows that the cosine to benign-mean cannot distinguish between two notions of “benign” because it measures alignment with a single direction rather than projecting against the full benign subspace This suggests that gate constructions summarizing benign reference statistics with a single mean direction are strictly weaker than those projecting against the full benign subspace The paper concludes that activation-cone defenses are structurally blind to prefilling at any single layer whether gated at prompt time or per generated token The response-time probing approach reduces prefilling ASR to ≤ 0. p. 05% on the canonical template family across all seven models at n = 40 This confirms the shallow alignment finding that safety training operates at the output level and can be bypassed by pre-filling compliance The unified architecture motivates an orthogonal split where AlphaSteer handles semantic attacks and response-halt handles prefilling

The paper's diagnostic protocols include a mechanism verification check using n=1000 instruction, template pairs per model to empirically verify assumed activation-geometric mechanisms The layer-depth-vs-generalization analysis is reusable for any depth-stratified probing claim The surface-feature reading discriminators include token frequency, benign query truncation, and template length tests to detect probes that fire on surface tokens rather than semantic features The library-version reproducibility check compares paper-shipped versus fresh-retrained halt rates under current versus published environments The paper's results show that the layer-depth landscape is model-specific, with Mistral having a usable semantic features depth at L=11 while Llama’s plateau may all be reading similar features This heterogeneity across models is consistent with the structural asymmetry of the activation-cone blind spot The findings confirm that a single-layer activation-cone gate fails on prefilling via whichever mechanism the prefill provides—geometric or distribution shift The paper's diagnostic protocols are designed to address standard evaluation gaps in current probing and steering safety work The deployment recommendation is that the unified system is the natural next step for Mistral where AlphaSteer wins on semantic attacks but is blind to prefilling

The dual-probe gated system achieves 0/40 on all models and reduces prefilling ASR to 8% on Mistral under the current environment The unified DSR of 0.983 on Mistral and 0.994 on Llama at n = 40 dominates both component paradigms The scaling to n = 200 shows the Llama dual-probe DSR reaching 0. p. 2990, while the Mistral unified fresh DSR reaches 0. p. 7940 This demonstrates that response-time probing realizes the contrapositive by classifying after the prefill commits to a particular trajectory The paper's results show that on Llama, the probe catches 80% of canonical prefilling attacks but only 65% of probe-aware adversarial variants This asymmetry is consistent with the layer-depth reading where the layer-1 probe catches variants whose window still carries compliance markers but not probe-aware search that specifically minimises that signal The paper's results show that on Mistral, the prefilling-specific probe does not fire on any adaptive variant This suggests that adaptive defense requires a general-purpose probe or the prompt-time component The paper's results show that the layer-selection criterion matters, and generalization-criterion selection is required over training-AUROC alone The paper's results show that the probe fires at chance on Gemma-4-31B under the stricter 30-token judge, reflecting its very low harm-rate prefilling distribution The paper's results show that the layer shift in Qwen-2.5 improves training-distribution detection but does not transfer to canonical-register novel-token generalisation The findings confirm that the two models diverge in their specific mechanism mix across six independent experiments The cumulative reading aligns with the corollary’s framing, where a single-layer activation-cone gate fails on prefilling via whichever mechanism the prefill provides The paper's diagnostic protocols are reusable beyond this paper for evaluating activation-level safety methods The deployment recommendation is that the unified system is the natural next step for Mistral where AlphaSteer wins on semantic attacks but is blind to prefilling

The paper's results show that on Llama, the dual-probe versus prompt-time paradigms compare DSR of 0.990 against AlphaSteer’s 0. p. 0698 The unified system on Llama reaches DSR 1. p. 6000 with the paper-pkl probe and 15% ASR under the Gemma-4-31B LLM judge The results show that on Mistral, the Unified fresh DSR is 0. p. 4940 while AlphaSteer is 75% This shows that the architectural composition motivated by the corollary yields DSR 0. p. 0983 on Mistral and 0.994 on Llama at n = 40 The paper's results show that the dual-probe system achieves DSR 0.706 on Mistral and 0. p. 2533 for the paper-pkl variant The findings confirm that a single-layer activation-cone gate fails on prefilling via whichever mechanism the prefill provides—geometric or distribution shift The paper's results show that the response probe fires above threshold τr = 0.5, and generation is halted and replaced with a refusal string otherwise The paper's results show that the response-probe FPR is 0% on 50 benign queries The paper's results show that the probe fires above threshold τr = 0.5, and generation is halted and replaced with a refusal string otherwise The paper's results show that the response-probe FPR is 0% on 50 benign queries The paper's results show that the unified system’s DSR is unaffected by the table’s drift footnote because it was measured in a current-env run The paper's results show that the Llama dual-probe DSR reaches 0.990 with n = 200 under the Gemma-4-31B LLM judge The paper's results show that the unified system on Llama reaches DSR 1.000 with the paper-pkl probe The paper's results show that the unified system on Mistral reaches DSR 1.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive defense architecture can be composed by combining AlphaSteer with response-time probe halt to achieve a unified DSR of 0.983 on Mistral and 0.994 on Llama at n=40, dominating individual paradigms while maintaining coverage against both prefilling and semantic attacks.

  2. Response-time probing can reduce prefilling ASR to less than 0/40 on all seven models at n=40, with the Llama result reaching 0/20 across runs when evaluated by a Gemma-4-31B LLM judge at n=200.

  3. The unified architecture achieves a DSR of 1.0 under the Gemma-4-31B LLM judge for both Mistral and Llama, resolving keyword judge discrepancies by identifying AlphaSteer's null-space intervention as pushing residual streams into incoherent space rather than producing actual harmful content.

  4. For Llama, the dual-probe system achieves a DSR of 0.990 at n=200 when evaluated by a Gemma-4-31B LLM judge, which outperforms AlphaSteer's score of 0.698 on the same metric and demonstrates that the probe catches 80% of canonical prefilling attacks but only 65% of probe-aware adversarial variants.

  5. Safety probing can be deployed at layer 14 for all steering injection, and the layer selection criterion should be based on generalization rather than training AUROC to ensure better robustness against novel templates.

Sources

Related papers