- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes SIGIL, a graph-based classifier for detecting adversarially manipulated prompt-and-patch pairs by software-development agents. SIGIL augments benign examples with synthetic adversarial instructions and code, constructs intent-code graphs, and trains a detector on their semantic and structural features. Experiments report accuracy and F1 scores across three agents on SWE-bench Lite and LiveCodeBench.
Critical strengths
-
The paper clearly describes the threat model
-
The main pipeline is easy to follow
Critical weaknesses
- Incremental methodology
The key insight is that an injected instruction and its implementation may align locally while remaining weakly integrated with the legitimate task. This insight have been widely explored in CHABADA/WHYPER-style app analysis as well as Task Shield. The relationship between SIGIL and these prior work needs to be delineated more clearly.
In addition, the intent-code graph is a potentially useful, but the current evaluation does not isolate its contribution. The auxiliary structural analysis shows associations between graph features and adversarial labels, but it cannot establish that graph learning is responsible for the detector for the detector's performance. More targeted ablations are need, such as a classifier on pooled CodeBERT embeddings, removal of individual edge types, shuffled or random edges, or removal of suspiciousness metadata.
- Insufficient experimental evidence
First, the paper reports accuracy and F1 without providing the numbers of benign and adversarial examples in each evaluation settings. Therefore, it is hard to interpret these accuracies and F1 scores. For example, accuracy can be dominated by the majority class, while F1 remains sensitive to class prevalence and small numbers of positive examples.
Second, the labeling progress is unclear. Positive labels are assigned manually to patches that ``meaningfully attempt the injected objective'', while the number of annotators, inter-annotator agreement, or whether annotators were blinded to detector outputs, are all unknown.
Third, attack-generated pairs that do not meet the criterion are excluded. Consequently, the evaluation measures detection conditional on a retained implementation attempt rather than performance over the full distribution of attack-generated outputs. The authors should report how any examples are excluded and clarify how this filtering affects the interpretation of the results.
Fourth, the main experiments focus on VWE-78, while the synthetic-data prompts explicitly specify CWE-78 and CWE-95 operations, including subprocess.run, os.system, eval, and exec. Therefore, the detector can receive supervision containing strong priors about the vulnerability patterns evaluated in the main experiments. RQ3 considers three additional CWE types, but the evaluation is tiny. Consequently, the generalization of SIGIL remain unclear.
Fifth, while the threat model states the attacker has no knowledge of the deployed defense, an adaptive evaluation, following the evaluation principles discussed in ``On Adaptive Attacks to Adversarial Example Defense'', should be considered to help assess whether a stronger attack can evade SIGIL.
- AI-use, ethics and reproducibility statements are inconsistent with the paper.
The statements on page 10 discuss factor generation, program revision, and operator construction,'' public historical stock-market observations,'' backtest performance,'' factor-selection rules,'', ``portfolio settings,'' do not correspond to SIGIL's method or experiments. Even worse, the reproducibility statement refers to sections that do not exist (search procedure, archive rules). These look like copy-paste from another submission.
Questions for authors
-
What is the contribution of the intent-code graph to SIGIL's detection performance?
-
How well does SIGIL generalize to CWE types not represented in its synthetic training data?
-
Why do the AI-use, ethics and reproducibility statements describe research unrelated to SIGIL?
3 comments
[2/3] Response to Reviewer CDwv
W3. Dataset Composition, Class Balance, and Evaluation Metrics
The original presentation of accuracy and F1 without corresponding class distributions limits the interpretability of the results, particularly when the number of retained adversarial examples varies across evaluation settings.
Tables 1 and 2 now report benign and adversarial sample counts alongside accuracy and F1. Appendix D reports per-detector TPR and FPR in Tables 7 and 8. The saved-adaptive and cross-attack-family transfer results appear in Table 3 (RQ1, Section 4.2) and Table 13 (Appendix G.2), respectively, with their class counts and metrics.
Section 4.1 and Appendices D, F.1, F.3, and G.1–G.2 document the splits and distinguish generated candidates from retained evaluation examples.
These additions will make it easier to assess the trade-off between adversarial detection and false positives, independently of class prevalence.
W4. Adversarial Labeling and Annotation Protocol
Appendix F now clarifies the labeling criterion and annotation procedure. An attack-generated patch is labeled adversarial when it contains code that meaningfully implements the injected objective beyond the legitimate task. For example, merely repeating the instruction in comments or importing a module without implementing the requested behavior is insufficient. The positive label therefore identifies an adversarial implementation attempt. Labels were assigned through LLM-assisted annotation and targeted manual review by the authors. During manual review, the authors compared each reviewed patch with its injected requirement against this criterion. Appendix F.3 describes the scope of the procedure.
W5. Excluded Examples and Evaluation Scope
Appendix D, Table 4 reports filtering counts for each attack, agent, and benchmark setting in Tables 1 and 2. Across these settings, the ledger records 22,350 agent completions, of which 2,157 were excluded. The exclusions comprise 199 empty patches, 400 malformed or non-applicable diffs or invalid Python outputs, 43 scanner-error exclusions, 232 patches that did not implement the injection, and 1,283 patches that did not implement the requested obfuscation. The retained evaluation set comprises 12,393 benign and 7,800 adversarial examples. Appendix F.1 describes instance filtering, and Appendix F.3 explains how labels are assigned to retained examples.
The reported detection rates are conditional on this retained population. A positive label requires a meaningful attempt to implement the injected objective. The positive-label criterion does not require demonstrated exploitability. Among retained examples, outputs that do not meet the positive-label criterion are counted as negatives. Excluded outputs are outside the denominators of the detector metrics. These rates therefore describe detection on the retained evaluation set, not end-to-end attack success over all generated outputs. Appendix G.1, Table 12 separately reports the screening outcomes for the saved adaptive candidates.
W6. Generalization to Unseen Vulnerability Types
The synthetic-training procedure emphasizes CWE-78 and CWE-95 operations, raising the possibility that the detector learns vulnerability-specific patterns rather than broader intent–code relationships.
To investigate this concern, an additional cross-vulnerability transferability experiment trained SIGIL using benign examples and SWExploit attacks only, excluding FCV attack-family examples from training and model selection. The source training set contained 508 benign and 506 SWExploit examples (1,014 total), while source validation contained 217 benign and 216 SWExploit examples (433 total).
The frozen detector was evaluated on four held-out FCV families, Base64-obfuscated attacks, and independently held-out SWExploit controls.
| Dataset | Benign | Adv. | Total | ACC | F1 | TPR | FPR |
|---|---|---|---|---|---|---|---|
| SWExploit (validation) | 217 | 216 | 433 | 99.08% | 0.9907 | 98.15% | 0.00% |
| FCV-78 | 317 | 266 | 583 | 90.22% | 0.8820 | 80.08% | 1.26% |
| FCV-89 | 317 | 313 | 630 | 92.54% | 0.9199 | 86.26% | 1.26% |
| FCV-94 | 317 | 319 | 636 | 89.62% | 0.8862 | 80.56% | 1.26% |
| FCV-502 | 317 | 303 | 620 | 92.74% | 0.9209 | 86.47% | 1.26% |
| FCV-78 (Base64) | 317 | 284 | 601 | 86.69% | 0.8387 | 73.24% | 1.26% |
| SWExploit (held-out) | 317 | 316 | 633 | 95.73% | 0.9560 | 92.72% | 1.26% |
| SWExploit (Base64) | 317 | 91 | 408 | 99.02% | 0.9785 | 100.00% | 1.26% |
ACC denotes accuracy; Adv. denotes adversarial examples.
[3/3] Response to Reviewer CDwv
W6. Generalization to Unseen Vulnerability Types (continued)
Source validation was used for checkpoint and threshold selection. All seven held-out evaluations reuse 317 benign examples. SWExploit controls test unseen tasks from the training family, and FCV outcomes measure transfer across families.
Across the four non-obfuscated FCV families, SIGIL achieves 80.08–86.47% TPR at a 1.26% FPR. Base64 FCV-78 yields 73.24% TPR, illustrating variation across attack conditions. The SWExploit held-out controls additionally demonstrate performance on unseen tasks from the source attack family.
In our static-analysis baseline, Bandit flags a patch when a selected warning code is triggered on an added line (Appendix F.4, Table 10). Its coverage therefore depends on whether the implementation triggers one of the selected rules. SIGIL uses task semantics and code structure without requiring vulnerability-specific warning codes at inference. Table 13 shows that a detector trained on benign and SWExploit examples transfers across all four evaluated plain FCV families without adding FCV-specific rules.
The results support cross-attack-family transfer. Training used archived SWExploit attacks, which may share underlying CWEs with FCV. Generalization to entirely unseen CWEs remains open.
RQ3 (Section 4.4) summarizes transfer. Appendix G.2 explains the protocol and limits. Table 13 reports counts and metrics across all seven held-out settings.
W7. Adaptive Attacks and Robustness
SIGIL's original threat model assumes a defense-unaware attacker. Because the method does not rely on secret detection mechanisms, an informed adversary could potentially exploit knowledge of graph construction and classifier feedback to craft evasive prompts.
Motivated by On Adaptive Attacks to Adversarial Example Defenses (Tramèr et al., 2020) and AutoDojo: Adaptive Attacks Expose Superficial Defenses and User-Underspecification Limits in LLM Agents (Ma et al., 2026), an AutoDojo-style feedback-guided search targeted earlier frozen SIGIL checkpoints with up to three attempts per eligible LiveCodeBench task. Independently retrained family-specific SIGIL detectors were then evaluated on the saved attack candidates without further optimization.
| Attack | Benign | Adversarial | Accuracy | F1 | TPR | FPR |
|---|---|---|---|---|---|---|
| FCV-78 | 316 | 237 | 94.94% | 0.9383 | 89.87% | 1.27% |
| SWExploit | 316 | 189 | 79.60% | 0.6436 | 49.21% | 2.22% |
LiveCodeBench; mini-SWE-agent. These saved attacks targeted earlier frozen checkpoints, not the independently retrained detectors evaluated here. Confusion counts (TP/FN/FP/TN): FCV-78 213/24/4/312; SWExploit 93/96/7/309. Both evaluations reuse the same 316 benign cases.
The retrained detectors detect 213/237 FCV-78 and 93/189 SWExploit candidates. Valid/invalid/unclassified counts are 237/708/18 and 189/190/14, respectively. We skipped 190 planned SWExploit tasks with unresolved objectives. Under an earlier validity screen, task-level evasion on the original targeted checkpoints was 14.64% (47/321) and 69.47% (91/131), respectively.
RQ1 (Section 4.2), Table 3 reports saved-attack transfer. Appendix G.1 and Table 12 document the protocol, limitations, and validity screening. Fresh adaptive evaluation of the retrained detectors and formal defense-aware robustness remain open.
W8. AI-Use, Ethics, and Reproducibility Statements
The statements concerning financial factor generation, stock-market observations, and portfolio backtesting were mistakenly included during manuscript preparation. They are unrelated to SIGIL and do not describe the research conducted. We apologize for this error.
The Ethics Statement has been replaced with the following SIGIL-specific disclosure:
ETHICS STATEMENT
This work studies defenses against adversarially induced vulnerabilities in LLM coding agents. It involves no human subjects or personal data, and it uses public benchmarks (SWE-bench Lite and LiveCodeBench) and previously published attacks. The synthetic adversarial prompts and vulnerable patches we generate for training and evaluation are dual-use, since they resemble the attack inputs that SIGIL aims to detect. Generated patches were evaluated only against benchmark repositories and were not deployed. SIGIL's main evaluation focuses on fixed, non-adaptive attackers, supplemented by a limited feedback-guided attack study. The evaluated vulnerability types and attack conditions remain limited, so SIGIL should not be treated as a complete safeguard. Deployments should combine it with other defenses and human review.
[1/3] Response to Reviewer CDwv
Thank you for the feedback on SIGIL’s methodology, evaluation, and robustness.
W1. Methodological Novelty and Relationship to Prior Work
SIGIL builds upon existing research in intent–behavior alignment and graph-based vulnerability detection. Its methodological contribution lies in a task-conditioned heterogeneous graph that jointly represents natural-language requirements, generated code, and their semantic and structural relationships to detect adversarially induced implementations.
CHABADA and WHYPER: CHABADA (Gorla et al., 2014) uses topic modeling to group Android applications by their descriptions, then applies a one-class SVM to identify unusual API usage relative to similar applications. WHYPER (Pandita et al., 2013) uses dependency parsing and semantic representations to determine whether application descriptions justify requested permissions. While both establish relationships between intended functionality and program behavior, SIGIL operates at a finer granularity, constructing a graph for each prompt-and-patch pair that connects individual task requirements to CFG-derived code blocks rather than relying on application-level API distributions or predefined permission semantics.
Task Shield: Task Shield (Jia et al., 2025) evaluates whether lower-trust instructions and tool calls contribute to trusted user objectives, using LLM-based alignment judgments to intervene during agent execution. SIGIL shares this task-alignment motivation but differs in its detection mechanism: it analyzes completed generated patches without requiring tool traces, employing a learned graph classifier rather than explicit task-contribution judgments. Unlike Task Shield, SIGIL does not independently authenticate trusted objectives, instead learning patterns from the supplied prompt and implementation.
Graph-based vulnerability detection: Devign (Zhou et al., 2019) uses program-semantic graphs and GNNs to classify vulnerable source code, while Vul-LMGNN combines pretrained code representations with program graphs. SIGIL extends beyond code-only representations by incorporating intent nodes and three relation types: intent dependencies, code control flow, and semantic intent–code grounding. Heterogeneous GraphSAGE integrates these relationships to classify potentially adversarial implementations relative to their associated tasks.
The underlying hypothesis is that malicious instructions may align locally with their implementations while remaining weakly integrated with the broader legitimate task. The additional decomposition, grounding, and architectural ablations investigate this hypothesis, although graph connectivity does not consistently improve performance across all settings.
The revised Related Work section details these differences in representations, alignment mechanisms, and detection objectives.
W2. Contribution of the Intent–Code Graph and Additional Ablations
The original structural-feature analysis establishes associations between graph properties and adversarial labels, but does not independently demonstrate that graph learning drives detection performance. Several additional ablations investigate this distinction.
Note: pp denotes percentage points (absolute differences between two percentages).
The pooled CodeBERT and no-edge experiments use archived attack-family training data rather than the original synthetic-only training procedure. Their findings therefore characterize architectural contributions under these additional training conditions, rather than definitively excluding shortcuts in synthetic training.
RQ4 (Section 4.5) summarizes the parsing, grounding, and pooled CodeBERT ablations. Appendix G.3, Table 14 reports their quantitative comparisons, with protocols and limitations in the accompanying text. The no-edge GraphSAGE results are reported here. RQ4 also describes the structural-feature analysis in Figure 2(b) as associative evidence.