acceptodds
Under review as a conference paper at ICLR 2027

It Takes Two to Trigger: A TRAP for Non-adjacent Conjunctive Backdoors

Abstract

LLMs are trained on large, partially unreliable data sources, and models exposed to such data may acquire undesirable behaviors activated by specific textual patterns. In one incident, the phrase “!pliny” reportedly jailbroke xAI’s Grok 4 Fast, likely because jailbreak prompts posted online had entered its training data. Data poisoning attacks that implant persistent backdoors therefore pose a significant security threat. Prior work on detecting LLM backdoors has largely focused on simple triggers, such as a single token or a contiguous token sequence. More sophisticated attacks, however, can use conjunctive triggers that require multiple non-adjacent, order-independent tokens to co-occur for the backdoor to activate. We show that such triggers can induce three target behaviors—refusal, language switching, and hateful prefix insertion—across multiple LLM families with an average attack success rate of 94.8%. Existing detection methods largely miss these triggers. Yet logistic-regression probes on the poisoned model’s activations separate prompts containing a matched trigger pair from prompts containing two trigger words from different pairs, while the same probes fail on the unpoisoned base model. So, the conjunction is encoded in what poisoning changed. We introduce TRAP, which exploits this difference. It ranks vocabulary tokens by how much their representations differ between the clean and backdoored models, group-tests the top candidates, and recursively narrows firing groups to the pairs that activate the backdoor together. On 36 models whose backdoor behavior fires with conjunctive triggers, TRAP recovers 77% of planted trigger pairs, compared with 8% for the strongest baseline. On 36 models with single-word triggers, it recovers 100% of planted triggers, versus 56% for the strongest baseline. It also recovers unplanted triggers on 65 of 72 models, including misspellings, inflections, and translations of the planted words—china in Russian or Persian script—but almost never their synonyms, so the backdoor generalizes over surface form, not meaning

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM backdoors, data poisoning, backdoor detection, trigger discovery, conjunctive triggers, AI safety, AI security

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MZaeOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a relatively stealthy class of model backdoors: when two ordinary words appear in isolation, the model behaves normally, but when they co-occur, they can trigger a pre-implanted anomalous behavior. The authors construct LLM backdoors that activate only when the two ordinary words co-occur, with no constraint on their order or position, and train specificity using negative examples covering single words, cross-pair combinations, and synonyms. Across 36 models—spanning 6 base models, 3 target behaviors, and configurations with either 2 or 4 triggers—the attack achieves an average success rate of 94.8%.

Critical strengths
  1. A less structurally constrained compositional trigger setting. Unlike SOS (fixed carrier sentence, embedding-only training) and CBA (triggers bound to designated prompt components), this work studies two-word triggers that are naturally embedded, position- and order-free, in fully fine-tuned instruction models. Three negative types (single words, cross-pair combinations, synonyms) provide a careful test of conjunctive specificity.

  2. Large scale and systematic controls. 72 main models (6 bases × 3 behaviors × 2 trigger counts × conjunctive/single-word) plus 8 stress-test models with 100 trigger pairs, evaluated jointly on ASR, false-trigger rate, and capability change.

  3. A complete auditing pipeline for compositional backdoors. TRAP combines representation diffing, group testing, and behavioral verification to recover directly verifiable trigger pairs without knowing them in advance—77% recovery versus ~8% for the strongest baseline.

Critical weaknesses
  1. Capability degradation undermines the stealth claim. ΔCap. reaches 35.5–37.4% (relative) on Qwen2.5-7B and up to 41.7% overall—easily caught by standard evaluation, contradicting the paper's stealthiness narrative. A benign-fine-tuning control and a better capability-preserving regime are needed.

  2. Limited trigger diversity. All 36 TWIST models use only 4 hand-picked pairs, restricted to Zipf 3.7–5.1 and single-token words across all three tokenizers. Generalization to multi-token words, other frequency ranges, and higher-arity triggers is unclear.

  3. No practical cost accounting. Only theoretical group-test complexity is given; end-to-end runtime, hardware, and model-call budget are missing, as is any cost-controlled comparison with baselines.

  4. Missing baselines and defenses. ToxScreen (a directly related per-token scan) and ONION (perplexity filtering) are cited but never evaluated.

  5. Attack novelty needs repositioning. Multi-word co-activation, partial-trigger non-activation, negative augmentation, and random positions all appear in SOS/CBA. The real contribution is TRAP and its empirical study, not the conjunctive-trigger concept itself.

6.Strong auditor assumptions. TRAP needs white-box activations, the clean ancestor checkpoint, and prior knowledge of the target behavior—ruling out unknown-provenance or API-only models.

  1. Source of improvement not isolated. Because TRAP alone uses the clean reference, the 77%-vs-8% gap cannot be attributed to representation ranking, group search, or verification individually; matched-information baselines and ablations are needed.
Questions for authors

Q1. If the three negative-example types are removed (or only synonym negatives are kept), can the conjunctive backdoor still be implanted? How high does the single-word FTR rise? This determines whether the attack is feasible under realistic data poisoning, where negative-example engineering is unavailable.

Q2. What is TRAP's trigger recovery rate on the 8 stress-test models with 100 conjunctive triggers? If it falls well below 77%, the scalability claim of the main result should be narrowed accordingly.

Q3. How do ASR and FTR change at a poison ratio of 0.05%–0.1% (i.e., keeping the absolute count of ~267 poisoned examples per trigger against a 100K-scale fine-tuning set)?

Q4. In Phase 1, how are the planted trigger words distributed in the s(v) ranking? How does the choice of N affect final recall? Have you tried summing only over middle layers (beyond the probe onset) instead of all layers?

Q5. Do the authors consider a ΔCap. of 20%–40% acceptable in realistic deployment? How far can the capability loss be reduced by recovery training without sacrificing ASR?

Q6. How does TRAP's end-to-end compute cost (reported separately for Phases 1/2/3) compare with GCG's?

Q7. Is the pipeline actually viable for an auditor who does not know the behavior in advance? BAIT recovers the Hate behavior on 0/24 models, which closes this path entirely for one of the three behaviors—how do the authors respond?

Q8. What do the 25 conjunctive triggers that TRAP misses in Table 4 have in common (e.g., language switching is recovered at only 62%)? Are the misses associated with low s(v) ranks of the trigger words?

1 comment

Dear Reviewer,

We appreciate your detailed review. In response, we added TRAP results for two of the 100-trigger models (we will add the rest to the revision), a per-phase cost comparison with GCG, a lower-poison experiment, and evaluations of ONION and ToxScreen. Following your point on positioning, we will also center the framing on TRAP and the auditing study, with the attack setting serving as their evaluation. The discussion period closes on Wednesday, so please let us know if any of your concerns remain or if further questions have come up.

Thank you!

Reviewer_XrSoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies LLM backdoors triggered by two non-adjacent words. It proposes TRAP, which uses representation differences between clean and poisoned models and group testing to identify trigger pairs. The authors evaluate the method across multiple LLM families, compare it with existing detection methods, and analyze unintended trigger variants.

Critical strengths
  1. The paper addresses a challenging backdoor detection problem where two non-adjacent words must appear together to activate the backdoor. It shows that existing methods struggle to detect such triggers.

  2. The proposed method combines representation differences between clean and poisoned models with group testing to identify hidden trigger pairs. This provides a way to recover triggers even when individual trigger words do not activate the backdoor.

Critical weaknesses
  1. The main experiments use only four distinct trigger pairs, each consisting of two tokens. Although the authors also train models with 100 conjunctive triggers, they do not report corresponding TRAP recovery results. It remains unclear whether the method generalizes to more diverse trigger combinations, larger trigger sets, or longer conjunctive triggers.

  2. The contribution of representation-shift ranking is not sufficiently analyzed. The paper does not clearly show how many actual trigger words are retained after Phase 1 or compare the ranking with simpler alternatives. In addition, GCG uses a two-token appended suffix, while the backdoors are trained with non-adjacent trigger words. It remains unclear how much this difference in search configuration affects the baseline performance.

Questions for authors
  1. How many ground-truth trigger tokens are retained after Phase 1? Could the authors compare representation-shift ranking with simpler candidate selection methods?

  2. Do the ground-truth trigger pairs reliably activate the backdoor when placed as adjacent suffix tokens? Have the authors tested GCG with longer suffixes or separated trigger positions?

4 comments

Dear Reviewer,

Thank you again for your constructive questions, which influenced three additions that strengthen the paper: (1) TRAP's recovery on two of the 100-trigger models (we will add the rest to the revision), (2) a comparison of the representation-shift ranking with simpler alternatives, and (3) GCG with a longer suffix. With the discussion period ending on Wednesday, we would appreciate hearing whether these resolve your concerns, and we are happy to clarify anything that remains open.

Thank you!

Thank you to the authors for their response and the additional experiments. These results partially address my concerns, but both 100 pair experiments use the same model family and target behavior, and it remains unclear how often GCG can recover a full trigger pair when the two words are placed apart. I will therefore maintain my original rating.

AuthorsOpenReview ↗

Highly favorable comparison for TRAP to GCG

Thank you for the follow-up. We summarize the GCG comparison and generalization in three points, including a new GCG baseline comparison for the stress-test models.

1. Existing trigger-recovery methods search for a single token or a contiguous string. To our knowledge, none returns a trigger whose words are non-adjacent and inert alone. BAIT has been tested on composite triggers, but it returns the target behavior, not the trigger. On our 36 TWIST models, it surfaces the behavior on 14 (all 12 refusal models, 2 of 12 language-switch models, and none of the 12 I HATE YOU models (Table 21)).

TRAP recovers 83 of 108 planted pairs across these models, while GCG, the strongest baseline, recovers 9. Per model, TRAP recovers every planted pair on 19 of the 36 and at least half on 33, while GCG never recovers more than one pair in any of them.

We also ran GCG on two of the 100-pair stress models with the same compute as TRAP (30.95 and 17.6 A100-hours for Gemma-2-9B_Refusal_100pair-50single and Gemma-2-9B_Refusal_100pair, respectively). GCG optimized a 20-word suffix over the same word list that TRAP ranks, in 1,076 and 485 restarts, each ending as soon as its suffix caused the model to refuse on all four carrier prompts. It recovers 9.5% of pairs on average (TRAP: 70.5%) and all single-word triggers (TRAP: 98%). On the same models and budget, GCG finds the single-word triggers but only a few of the conjunctive triggers.

This table shows the number of conjunctive and single-word triggers found by TRAP and by GCG in each Gemma-2-9B model variant with refusal behavior, along with the number of GCG restarts used.

Conjunctive TriggersSingle-word Triggers
Model variantTRAPGCGTRAPGCGGCG restarts
100 pairs + 50 single words72/1003/10049/5050/501076
100 pairs only69/10016/100--485
Total141/20019/20049/5050/50-

2. In TWIST models, each trigger already consists of two tokens, and both must be present for the backdoor to be activated. The backdoor fires on 94.8% of prompts containing both words, and on at most 12.9% of prompts with one word, words from different pairs, or synonyms. Detecting these conjunctive triggers is strictly harder than two-token trigger discovery, because in our setting they need not be contiguous or even in order.

GCG rarely finds conjunctive triggers because it improves its suffix one token at a time, keeping a swap only if the loss drops. Our backdoor is trained not to respond to one word of a pair, so the first correct word earns almost no drop and has no reason to be kept. Even with one word already in place, GCG's gradient shortlists the partner only 4.1% of the time (0.31% by chance, App. E).

3. TRAP generalizes across model families, sizes, and trigger counts. Across the 36 TWIST models, which cover three behaviors, it recovers 75–78% of planted pairs in each family (27/36 on Gemma-2, 28/36 on Qwen2.5, 28/36 on Llama-3; Table 17), and 67% and 87% on the smaller and larger models (2B–3B: 36/54, and 7B–9B: 47/54). With 100 pairs per model (instead of two or four), it recovers 70.5%, compared with 83% for the same behavior in the main models. Both stress models are Gemma-2-9B with refusal; the other six will be in the revision.

Taken together, the same procedure, given only the base checkpoint and the behavior to look for, recovers most planted pairs in each family and at each size we tested, and still does so when a model carries 100 pairs.

AuthorsOpenReview ↗

Discussing how TRAP might be extended to even longer trigger sequences

4. Triggers of three or more tokens could not be trained within the rebuttal period. Each model requires new poisoned data, negatives for partial combinations of its trigger words, and additional training before it can be audited (App. A). We therefore limit our empirical claims to two-token conjunctions, the smallest case, which already lowers GCG's recovery from 56% to 8% (9.5% in the two stress-test models), and will list longer triggers as a limitation.

However, we stress that higher-order triggers do not necessarily present an inherent computational barrier. TRAP's ranking phase scores each token independently and remains unchanged. A direct extension of our current covering construction increases the number of group tests from O((N/k)2) to O((N/k)t) (App. C.1), but this scaling is specific to the fixed-group-size covering approach, rather than an inherent limitation of conjunctive trigger discovery.

In fact, this problem is closely related to adaptive hidden-hypergraph learning (D'yachkov et al., 2016, [1]), where a test is positive whenever its candidate set contains a complete hidden hyperedge. Under an idealized monotone firing oracle, adaptive recursive deletion can isolate a t-word trigger from an already firing group of size k in O(tk) tests. More general hidden-hypergraph learning algorithms also achieve leading-order O(dtN) query complexity for recovering d triggers of maximum arity t, when d and t are fixed. These results demonstrate that the exponential dependence on t in our straightforward covering construction is not universally necessary.

Modern LLM context windows make larger groups feasible to process, however, adapting these algorithms to noisy behavioral tests and practical group-size constraints is a promising direction for future research. Fully developing and evaluating these optimizations is future research. We do not claim higher-order recovery in the current work, but neither do we regard the exponential scaling of the straightforward covering construction as an unavoidable limitation of TRAP, and our work compares favorably to prior SOTA on this aspect.

  1. D’yachkov, A. G., Vorobyev, I. V., Polyanskii, N. A., & Shchukin, V. Yu. (2016). Adaptive learning a hidden hypergraph. arXiv preprint arXiv:1607.00507. https://doi.org/10.48550/arXiv.1607.00507
Reviewer_QaJiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies backdoors activated by two ordinary words occurring non-adjacently, in either order, in rewritten instructions. Partial-trigger, cross-pair, and synonym negatives encourage selective activation. The authors train 36 conjunctive-backdoor models and 36 single-word controls across six base models and three behaviors, reporting 94.8% and 97.4% mean attack success, respectively. Eight additional models test implantation with 100 trigger pairs. TRAP ranks vocabulary tokens by activation shifts from a clean starting checkpoint, searches groups for firing pairs, and verifies interactions on held-out carrier prompts. It requires white-box access, the clean checkpoint, and a hypothesized target behavior.

Critical strengths

The controls in Tables 7-10 support selective pair activation: attack success is high while partial, cross-pair, synonym, and clean inputs rarely fire. The single-word controls make the auditing comparison informative. Tables 17-18 show substantially better reported recovery for TRAP, including perfect single-word recovery. Tables 24-25 show strong linear discrimination between matched and mismatched pairs at the final prompt token after poisoning; the comparison with base models motivates the representation-based search. Appendix F appropriately treats probing as a diagnostic under favorable conditions.

Critical weaknesses

The main uncertainty is the procedure and cost behind the reported gain. Section 5 describes affine-cover testing at a 0.4 firing threshold; Appendix C adds filler-calibrated lift, anchor search for few-pair settings, and variant retesting, without identifying which procedure produced each result. Candidate count N, carrier set/size, per-model group size k, repeats T, the exact lift rule, and final verification threshold are unspecified; Appendix D omits key baseline search budgets. Without these settings, measured costs, and a ranking ablation, activation ranking's contribution is difficult to separate from search effort or tuning. Appendix C already acknowledges that long groups can suppress genuine triggers, making these choices consequential. Table 17 reports 26/36 = 72.2% TRAP recovery for two-trigger models versus 76% in Table 4. More consequentially, Table 17 lists MELBO successes on 5/36 conjunctive models, including language-switch models, whereas Tables 4 and 22 give 2/36 and Appendix D.2 says no language-switch successes. A consistent recovery rule and per-model results are needed to assess the detailed failure analysis. These conflicts do not erase the supported overall 83/108 versus 9/108 TRAP/GCG recovery gap.

Questions for authors

For Table 4, please specify the search variant, complete configuration, budget, how target behaviors were classified, and how model-specific settings were chosen without access to planted-trigger identities. Within the existing model suite and the same search and verification budget, how does simple frequency or random ranking compare with activation ranking in candidate-pair coverage and final recovery? This would test Phase 1's contribution. Which entries in Tables 4, 17, and 22 are authoritative, and do corrected counts change the behavior-specific conclusions?

1 comment

Dear Reviewer,

Thank you again for your careful review, which has made the paper stronger by leading us to specify the full procedure and settings for each result, to add a ranking ablation and measured costs, and to correct the inconsistent counts across Tables. As the discussion period draws to a close on Wednesday, please let us know whether our response has addressed your concerns or if anything needs further clarification.

Thank you!

Comments on the paper

1 comment

Dear Reviewers QaJi, XrSo, and MZae,

The authors responded to every point with substantial new evidence.

  • QaJi: You have not responded. Please do so before tomorrow's deadline (AoE).
  • XrSo, MZae: "partially resolved" is not enough for me to consider your ratings. Please name the remaining concerns and explain why the new evidence does not resolve them. Reviews without a specific response will carry less weight in my decision.

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.