Pith. sign in

REVIEW 3 major objections 4 minor 59 references

CyberForge: Verified Vulnerability Injection at Repository Level for Cybersecurity Agent Training

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read CyberForge claims that execution-validated synthetic vulnerabilities can replace CVE mining as training data, improving repair agents by up to 14.7 points and transferring to other languages.

desk verdict Solid, well-engineered corpus paper; the training-data claim mostly holds, but the SEC-bench generalization claim needs an explicit project-overlap check. read the letter →

arxiv 2608.06471 v1 pith:MN35A37F submitted 2026-08-06 cs.CR cs.AIcs.SE

Amine Lbath , Manan Suri , Aurelien Delaitre , Vadim Okun , Massih-Reza Amini , Ram D. Sriram , Dinesh Manocha This is my paper · ORCID
classification cs.CRcs.AIcs.SE
keywords vulnerabilityinjectionproof-of-vulnerabilitysynthetictrainingdataLLMagentfine-tuningpatchrepairC/C++securitydifferentialvalidationbenchmarks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CyberForge's central claim is that security training data for defensive agents does not have to be mined from disclosed vulnerabilities: it can be generated by injecting weaknesses into real C/C++ projects and keeping only instances that are validated by execution. Each accepted instance must compile, pass the project's own unit tests on both the original and injected build, and come with a proof-of-vulnerability that triggers only on the injected build. On this corpus the paper collected agent trajectories from two teachers and fine-tuned three open-weight student models, and all six configurations improved patch repair on the SEC-bench task by +3.3 to +14.7 points, with the largest student reaching 72.7% against its teacher's 74.0%. The same training improved repair on PatchEval, a benchmark in Go, JavaScript, and Python, despite a C/C++ training corpus. If the claim holds, it removes the main bottleneck limiting security-agent training: scarcity of reproducible, executable vulnerability data.

What carries the argument

The load-bearing mechanism is a differential proof-of-vulnerability (PoV) oracle, applied to every candidate instance. It enforces two conditions: the injected build must pass the project's existing unit tests, and a generated PoV must trigger on the injected build while leaving the clean build unaffected. This oracle is shared by two complementary generation pipelines: a fuzzer-guided pipeline that targets functions reachable by existing harnesses, and an agentic in-context pipeline that selects sites by static analysis, retrieval of similar historical code patterns, and autonomous exploration. The oracle turns a synthetic bug into a certified training signal: it guarantees the injected project is still buildable and test-passing, that the weakness is latent under normal execution, and that it is demonstrably reachable by an adversarial input. The paper also uses the oracle to decide which teacher trajectories to keep, discarding any trajectory whose claimed success is not verified.

What would settle it

Compute the project-level overlap between the 80 training projects and the 24 SEC-bench projects (and any PatchEval repositories), retrain after deleting all training instances from overlapping projects, and check whether the SEC-bench and PatchEval improvements survive; if the gains collapse on fully disjoint projects, the generalization claim is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that vulnerability injection with differential validation is a viable and scalable substitute for CVE mining. The core result is empirical: fine-tuning on oracle-verified trajectories over 1,034 injected vulnerabilities, across 80 projects and 63 weakness categories, improves a repair agent's success rate on the SEC-bench benchmark in every configuration tried. The improvements are not confined to the training distribution; every configuration also improves on PatchEval, whose Go, JavaScript, and Python instances are absent from the C/C++ corpus. The paper also claims the injected edits resemble real CVE patches in locality, with a two-sample Kolmogorov–Smirnov distance of 0.165 against a 0.190 real-versus-real baseline. The mechanism it credits is not better injection alone but the validation machinery: in the agentic pipeline, a naive single pass compiles and passes tests in 68.2% of attempts but validates 0% of them, while the full workflow raises validated yield to 7.5%.

Load-bearing premise

The evaluation benchmarks and the training projects may share source-level ancestry, so the reported gains could partly reflect memorization of familiar codebases rather than a general vulnerability-repair skill.

Editorial extensions

If this is right

  • Corpus growth no longer depends on the rate of public vulnerability disclosure; any buildable project with a fuzzing harness can supply new training instances.
  • Smaller students gain the most: the 12B model roughly doubles its SEC-bench score, and self-distillation improves every scale, so even a strong model can learn from its own verified trajectories.
  • Fine-tuning transfers repair behavior, not just scores: students adopt the teacher's edit-and-verify workflow, with verification coverage rising from 20.7% to 82.7% at 12B and format-error rates collapsing.
  • The gains carry to out-of-distribution languages (Go, JavaScript, Python), suggesting the corpus teaches general vulnerability-repair concepts rather than encoding C/C++ syntax patterns.
  • Ensembling two students trained under different teachers beats either single student and even beats the stronger teacher (82.0% vs 74.0% at 31B), a post-hoc result the paper computes from the same runs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension not explored in the paper is to measure project-level overlap between the training projects and evaluation projects; if gains persist after excluding any overlapping repositories, that would strengthen the generalization claim beyond what is currently reported.
  • Because validation is execution-based, the same differential oracle could be reused as a filter or reward signal for other synthetic-data generators, including those targeting other languages or non-memory-safety bug classes.
  • The paper's edit-locality comparison suggests a cheap quality probe: monitor the Kolmogorov–Smirnov distance between injected and real patches over time as the pipeline scales; if distance grows, the corpus may drift toward idiosyncratic, easily caught bugs.
  • One could test whether the trajectories' benefit comes from the successful-patch examples or from the failed attempts discarded by the oracle; training on filtered failures alone might reveal whether negative signal matters.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper presents CyberForge, a framework that generates repository-level vulnerability-injection training data for cybersecurity agents. CyberForge injects weaknesses into real C/C++ OSS-Fuzz projects, requires each injected build to pass the project's unit tests, and validates each instance by a differential proof-of-vulnerability oracle (PoV triggers on the injected build but not on the clean one). The released corpus contains 1034 validated vulnerabilities across 80 projects and 63 CWE categories. The authors collect agent trajectories on these instances using two teachers (GPT-5.4-mini and Gemma 4 31B) and fine-tune three Gemma 4 students (E4B, 12B, 31B). They report that all six student-teacher configurations improve SEC-bench patch repair by +3.3 to +14.7 percentage points, with the strongest 31B student reaching 72.7% against its teacher's 74.0%, and that all configurations also improve on PatchEval, a Go/JavaScript/Python benchmark disjoint in language from the C/C++ training corpus. The paper also claims that the injected edits have edit-locality statistics similar to real CVE patches under a real-versus-real noise floor.

Significance. If the empirical claims hold, this is a useful and timely contribution: it decouples security training-data generation from the rate of public CVE disclosure, provides reproducible containerized environments, validates every instance with an executable differential oracle rather than static heuristics, and shows consistent downstream gains across model scales and teachers. The cross-language PatchEval results and the self-distillation results are valuable pieces of evidence. The paper is also transparent about pipeline costs and reports a detailed cost accounting. The two main open issues are (1) unquantified project-level overlap between the CyberForge OSS-Fuzz training projects and the SEC-bench evaluation instances, which is load-bearing for the headline in-domain claim, and (2) the absence of repeated-seed or significance information behind the 'all six configurations improve' statement. These issues are fixable and do not undermine the framework's potential, but they must be resolved before the central claims can be accepted at face value.

major comments (3)
  1. [§3.1, Appendix A.4, Table 5] The paper never reports the project-level overlap between the 80 CyberForge training projects listed in Table 5 and the 24 SEC-bench projects. Because CyberForge deliberately targets OSS-Fuzz projects (§3.1) and Appendix A.4 states that SEC-bench's 300 real instances are split into '200 from the cve split and 100 from the oss split,' the two corpora plausibly share several projects (for example brotli, libarchive, libpng, or openssl). If common projects exist, the SEC-bench repair gains of +3.3 to +14.7 points (§5.1, Table 2) could reflect project-specific familiarity (file layout, build conventions, format-specific input quirks) rather than a general vulnerability-repair capability. This is load-bearing for the central claim that the gains are not memorization. Please report the exact project-level and instance-level overlap and re-evaluate on the disjoint subset of SEC-bench; PatchEval provides partial transfer evidence, but it does not validate the headline in-domain numbers on a disjoint C/C++ set.
  2. [§5.1, Table 2, Appendix D] Every fine-tuned model is trained once with seed 42 (Table 13) and evaluated on fixed benchmark sets, so there is no variance information and no significance testing behind the statement that all six student-teacher configurations improve. The smallest reported gains (+3.3 and +4.7 percentage points at E4B) are within the range that a single training run could flip under a different initialization or trajectory-sampling order, particularly given the small absolute number of solved instances (the E4B base solves only 9 of 150 SEC-bench instances). Please report multiple seeds (or permutation/bootstrap intervals over benchmark instances) and show that the improvements hold with accompanying uncertainty intervals.
  3. [§5.2 and Appendix A.4] The 'near-realistic' edit-locality claim is asserted from a single KS distance on functions changed (0.165 versus a real-to-real baseline of 0.190), but the same appendix reports injected-to-real distances of 0.123 for files touched and 0.243 for hunks, against baselines of 0.065 and 0.205. On the files-touched axis the injected corpus is nearly twice as far from real patches as two real corpora are from each other, so the statement that 'injected edits broadly resemble real patches' is not uniformly supported by the reported numbers. Please report all five axes with uncertainty estimates and either temper the realism claim or provide a multivariate assessment.
minor comments (4)
  1. [Figure 4] The pie-chart slices are described with percentages in the text but the figure itself does not show the values; adding percentage labels would make the failure-stage analysis self-contained.
  2. [Table 5] The caption says 'All 100 qualified OSS-Fuzz projects' but several rows show zero P1/P2 instances and zero CWE categories; please distinguish clearly between qualified projects and projects that contributed at least one validated instance, or mark the non-contributing rows explicitly.
  3. [Appendix E.5] The release metadata uses the value 'fuzz_poc_guided' as a producer identifier, while the paper consistently calls this pipeline 'fuzzer-guided'; please standardize the terminology between the paper and the released corpus metadata.
  4. [Listing 1] The text says the deleted line is shown in red, but the listing is monochrome; please use an explicit deletion marker or a footnote so the removed line is unambiguous in print.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CyberForge training and SEC-bench/PatchEval evaluation are connected empirically, not by construction.

full rationale

The claimed derivation chain is self-contained and non-circular. CyberForge builds a synthetic vulnerability corpus by injecting defects into OSS-Fuzz projects and admits instances only through a differential PoV oracle (unit tests pass; PoV fires on the injected build only). The downstream claim is that fine-tuning on teacher trajectories over this corpus raises SEC-bench and PatchEval repair scores. Nothing in that chain equates the training objective to the evaluation outcome: no SEC-bench instance or label enters the corpus, no parameter is fit to SEC-bench scores, and the differential oracle is not SEC-bench's oracle. The closest passage is Section 3.4: "Task construction matches SEC-bench (Lee et al. 2025), following the same methodology as (Luo et al. 2026), so that training and evaluation align," together with the trajectory-environment normalization in Appendix D. That creates a legitimate in-distribution/possible-contamination concern (project overlap between the 80 OSS-Fuzz training projects and SEC-bench's projects is not reported), but this is a validity or reproducibility risk, not circularity: the SEC-bench numbers are not implied by construction from the training data. The KS-distance realism comparison (Section 5.2, Appendix A.4) is a descriptive statistic against SEC-bench, not an optimized target. The only author self-citation, AVIATOR (Lbath et al. 2026), is used as comparative related work and is not load-bearing. No uniqueness theorem or ansatz is imported from the authors' prior work.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a small set of domain assumptions: the differential oracle correctly identifies true vulnerabilities, OSS-Fuzz projects represent realistic targets, and the edit-locality KS statistic is a meaningful realism metric. The pipeline also contains hand-chosen weights and retry/time budgets that shape the corpus but are not fitted to the evaluation outcome.

free parameters (2)
  • Candidate scoring weights
    Pipeline 1 Step 2 (Candidatescoringandselection) ranks functions using a 'weighted combination' of a function score and a triggerability score; the weights are not specified, so the composition of the corpus depends on hand-chosen weights.
  • Pipeline retry limits and time budgets
    Section 3.2 and Appendix B set retry limits for PoV and injection (5 and 3), a 90-second post-hoc fuzz run, and a 60-second max_total_time in the PoV driver. These hand-set values shape which instances are validated and the teacher trajectory collection, without being fitted to the downstream benchmark.
assumptions (3)
  • domain assumption The differential PoV oracle is a sufficient certificate that an injected change is a real vulnerability.
    Section 3.3 accepts an instance if the injected build passes unit tests and the PoV triggers on it but not the clean build. This assumes the oracle has no false positives (e.g., sanitizer findings unrelated to the injection) and that the resulting bugs are security-relevant.
  • domain assumption OSS-Fuzz projects and their unit test suites are representative of real-world C/C++ software.
    Section 3.1 builds the entire corpus from OSS-Fuzz projects that pass their own tests at 100% across five runs. The paper's realism claims and downstream generalization assume this population of well-tested open-source projects stands in for real-world code.
  • domain assumption KS distance between edit-size distributions is a meaningful proxy for task-relevant realism of injected vulnerabilities.
    Section 5.2 and Appendix A.4 use a two-sample KS distance of 0.165 against a real-to-real baseline of 0.190 to claim 'near-realistic' edits. This assumes that similarity along functions/files/hunks/changed-lines/added-lines captures the properties that make a vulnerability useful for repair training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CyberForge: Verified Vulnerability Injection at Repository Level for Cybersecurity Agent Training." pith.science (2026). https://pith.science/paper/MN35A37F

@misc{pith2026260806471,
  author       = {Pith},
  title        = {Pith review of: CyberForge: Verified Vulnerability Injection at Repository Level for Cybersecurity Agent Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MN35A37F}},
  note         = {Machine review of arXiv:2608.06471}
}
read the original abstract

Despite recent advances, frontier large language model (LLM) agents remain limited in discovering and patching complex vulnerabilities in real-world software. Generally available agents can already aid attackers, who only need to find one exploitable weakness, while defenders must continuously identify and patch all vulnerabilities across fast-growing codebases. Stronger defensive agents would help close this gap, yet the scarcity of security training data with reproducible build and execution environments remains a bottleneck. We present CyberForge, a framework that synthesizes executable, repository-level security training data by injecting vulnerabilities into real C/C++ projects. It validates each instance dynamically: the injected build must pass the project's unit tests, and generated proof-of-vulnerability (PoV) must trigger on the injected build and not on the clean one. CyberForge is not limited by the availability of disclosed vulnerabilities, therefore it can scale in comparison to data augmentation techniques which rely on historic CVE data. The resulting corpus holds 1034 validated vulnerabilities across 80 projects and 63 weakness categories, with edit locality similar to real CVE patches under a real-versus-real noise floor. Fine-tuning on trajectories collected over this corpus improves SEC-bench patch repair by +3.3 to +14.7 points, in all six configurations of three model scales and two teachers, with the 31B student reaching its GPT-5.4-mini teacher, 72.7% against 74.0%. These gains generalize out of distribution to PatchEval, a corpus containing other programming languages, where every configuration also improves and the 31B student passes its teacher.

Figures

Figures reproduced from arXiv: 2608.06471 by the authors.

Figure 1
Figure 1. Overview of CyberForge. Two injection pipelines synthesize candidate vulnerable/patch pairs from OSS-Fuzz C/C++ [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Training data scaling evaluated on SEC-bench. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Analysis of teacher transfer, and base retention on [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 6
Figure 6. Figure 6: Visualization of behavorial metrics comparing the [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 5
Figure 5. Figure 5: Analysis of coarse agent action behavior over 30 turns, demonstrating transfer of agent action density from teacher [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: What the corpus covers, by pipeline. (a) The 20 most frequent CWE identifiers; Table 7 gives their names. (b) The 20 [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: P1: CyberForge’s fuzzer-guided injection pipeline. Vulnerability Generation Exploit Validation Stage 1: Input Stage 2: Injection Stage 3: Reachability Stage 4: Exploit Stage 5: Verifier Selected Location Selected CWE Call chain Dataflow Reachability tier from LLM or re…
Figure 9
Figure 9. Figure 9: P2: CyberForge’s agentic in-context injection pipeline. ness at a selected location, using compiler and unit-test feed￾back. Once the project builds and its tests pass, a taint￾analysis agent examines CodeQL-extracted dataflow from the source (an entry point through wh…
Figure 10
Figure 10. Figure 10: Time and cost per teacher trajectory generation run. Duration is read on the lower axis, cost on the upper one. Boxes [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Time and cost breakdown of the P1 fuzzer-guided injection pipeline per attempt. The injection agent is the only stage [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Time and cost per sub-agent of the P2 agentic injection pipeline per attempt. [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 47 canonical work pages

  1. [1]

    2026 , howpublished =

    Claude Mythos Preview , author =. 2026 , howpublished =

    work page 2026

  2. [2]

    2024 , howpublished =

    Project Naptime / Big Sleep: From Naptime to Big Sleep---Evaluating Offensive Security Capabilities of Large Language Models , author =. 2024 , howpublished =

    work page 2024

  3. [3]

    How I Used o3 to Find

    Heelan, Sean , year =. How I Used o3 to Find

    work page

  4. [4]

    2024 , url =

    Fang, Richard and Bindu, Rohan and Gupta, Akul and Kang, Daniel , journal =. 2024 , url =

    work page 2024

  5. [5]

    Teams of

    Fang, Richard and Bindu, Rohan and Gupta, Akul and Zhu, Qiusi and Kang, Daniel , journal =. Teams of. 2024 , url =

    work page 2024

  6. [6]

    Asleep at the Keyboard?

    Pearce, Hammond and Ahmad, Baleegh and Tan, Benjamin and Dolan-Gavitt, Brendan and Karri, Ramesh , booktitle =. Asleep at the Keyboard?. 2022 , doi =

    work page 2022

  7. [7]

    2025 , url =

    Shahid, Muhammad Mohsin and others , journal =. 2025 , url =

    work page 2025

  8. [8]

    2024 , url=

    John Yang and Carlos E Jimenez and Alexander Wettig and Kilian Lieret and Shunyu Yao and Karthik R Narasimhan and Ofir Press , booktitle=. 2024 , url=

    2024

Show all 59 references
  1. [9]

    2024 , url =

    Ding, Yangruibo and Fu, Yanjun and Ibrahim, Omniyyah and Sitawarin, Chawin and Chen, Xinyun and Tizpaz-Niari, Saeid and Millstein, Todd and Ray, Baishakhi and Wang, Yizheng , booktitle =. 2024 , url =

    2024

  2. [10]

    and Yang, John and Wettig, Alexander and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , booktitle =

    Jimenez, Carlos E. and Yang, John and Wettig, Alexander and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , booktitle =. 2024 , url =

    2024

  3. [11]

    Training Software Engineering Agents and Verifiers with

    Pan, Jiayi and Wang, Xingyao and Neubig, Graham and Jaitly, Navdeep and Ji, Heng and Suhr, Alane and Zhang, Yizhen , booktitle =. Training Software Engineering Agents and Verifiers with. 2025 , note =

    2025

  4. [12]

    and Wettig, Alexander and Khandpur, Kabir and Zhang, Yanzhe and Hui, Binyuan and Press, Ofir and Schmidt, Ludwig and Yang, Diyi , booktitle =

    Yang, John and Lieret, Kilian and Jimenez, Carlos E. and Wettig, Alexander and Khandpur, Kabir and Zhang, Yanzhe and Hui, Binyuan and Press, Ofir and Schmidt, Ludwig and Yang, Diyi , booktitle =. 2025 , note =

    2025

  5. [13]

    2025 , note =

    Jain, Naman and others , booktitle =. 2025 , note =

    2025

  6. [14]

    , booktitle =

    Wei, Yuxiang and Duchenne, Olivier and Copet, Jade and Carbonneaux, Quentin and Zhang, Lingming and Fried, Daniel and Synnaeve, Gabriel and Singh, Rishabh and Wang, Sida I. , booktitle =. 2025 , note =

    2025

  7. [15]

    2025 , url =

    Lee, Hwiwon and others , journal =. 2025 , url =

    2025

  8. [16]

    and others , booktitle =

    Zhang, Andy K. and others , booktitle =. 2025 , note =

    2025

  9. [17]

    2025 , note =

    Zhu, Yuxuan and others , booktitle =. 2025 , note =

    2025

  10. [18]

    2026 , eprint=

    CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale , author=. 2026 , eprint=

    2026

  11. [19]

    arXiv preprint arXiv:2408.08926 , year =

    Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models , author =. arXiv preprint arXiv:2408.08926 , year =

    arXiv

  12. [20]

    Shao, Minghao and Jancheska, Sofija and Udeshi, Meet and Dolan-Gavitt, Brendan and Xi, Haoran and Milner, Kimberly and Chen, Boyuan and Yin, Max and Garg, Siddharth and Krishnamurthy, Prashanth and Khorrami, Farshad and Karri, Ramesh and Shafique, Mohammad , booktitle =

  13. [21]

    arXiv preprint arXiv:2508.00910 , year =

    Cyber-Zero: Training Cybersecurity Agents without Runtime , author =. arXiv preprint arXiv:2508.00910 , year =

    arXiv

  14. [22]

    arXiv preprint arXiv:2508.18370 , year =

    Training Language Model Agents to Find Vulnerabilities with CTF-Dojo , author =. arXiv preprint arXiv:2508.18370 , year =

  15. [23]

    2023 , doi =

    Nong, Yu and Ou, Yuzhe and Pradel, Michael and Chen, Feng and Cai, Haipeng , booktitle =. 2023 , doi =

    2023

  16. [24]

    2023 , url =

    Nong, Yu and Fang, Richard and Yi, Guangbei and Zhao, Kunsong and Luo, Xiapu and Chen, Feng and Cai, Haipeng , journal =. 2023 , url =

    2023

  17. [25]

    2025 , note =

    Xu, Xiangzhe and Su, Zian and Guo, Jinyao and Zhang, Kaiyuan and Wang, Zhenting and Zhang, Xiangyu , booktitle =. 2025 , note =

    2025

  18. [26]

    From CVE Entries to Verifiable Exploits: An Automated Multi-Agent Framework for Reproducing

    Saad Ullah and Praneeth Balasubramanian and Wenbo Guo and Amanda Burnett and Hammond Pearce and Christopher Kruegel and Giovanni Vigna and Gianluca Stringhini , journal =. From CVE Entries to Verifiable Exploits: An Automated Multi-Agent Framework for Reproducing. 2025 , url =

    2025

  19. [27]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Devign: Effective Vulnerability Identification by Learning Comprehensive Program Semantics via Graph Neural Networks , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  20. [28]

    2023 , url =

    Chen, Yizheng and Ding, Zhoujie and Lamotte, Wyatt and Bhatt, Dhruv , booktitle =. 2023 , url =

    2023

  21. [29]

    ACM Transactions on Software Engineering and Methodology , volume =

    Large Language Model for Vulnerability Detection and Repair: Literature Review and the Road Ahead , author =. ACM Transactions on Software Engineering and Methodology , volume =. 2025 , doi =

    2025

  22. [30]

    Fixing Security Vulnerabilities with

    Zhang, Yuntong and others , booktitle =. Fixing Security Vulnerabilities with. 2026 , note =

    2026

  23. [31]

    Disrupting the First Reported

    Anthropic , year =. Disrupting the First Reported

  24. [32]

    2026 , url =

    Wang, Zhun and Schiller, Nico and Li, Hongwei and He, Jingxuan and Holz, Thorsten and Song, Dawn , journal =. 2026 , url =

    2026

  25. [33]

    Shi, Tianneng and Rheem, Robin and Jiang, Dongwei and Wang, Mona and Guo, Wenbo and Song, Dawn , journal =

  26. [34]

    2026 , url =

    Lbath, Amine and Amini, Massih-Reza and Delaitre, Aurelien and Okun, Vadim , journal =. 2026 , url =

    2026

  27. [35]

    2024 , doi =

    Daneshvar, Seyed Shayan and Nong, Yu and Yang, Xu and Wang, Shaowei and Cai, Haipeng , journal =. 2024 , doi =

    2024

  28. [36]

    2025 , url =

    Yu, Yao-Ching and Chiang, Tsun-Han and Tsai, Cheng-Wei and Huang, Chien-Ming and Tsao, Wen-Kwang , journal =. 2025 , url =

    2025

  29. [37]

    Frontier

    Potter, Yujin and Guo, Wenbo and Wang, Zhun and Shi, Tianneng and Li, Hongwei and Zhang, Andy and Kelley, Patrick Gage and Thomas, Kurt and Song, Dawn , journal =. Frontier. 2025 , url =

    2025

  30. [38]

    To Defend Against Cyber Attacks, We Must Teach

    Zhuo, Terry Yue and Ding, Yangruibo and Guo, Wenbo and Meng, Ruijie , journal =. To Defend Against Cyber Attacks, We Must Teach. 2026 , url =

    2026

  31. [39]

    2026 , howpublished =

    Scaling Trusted Access for Cyber with. 2026 , howpublished =

    2026

  32. [40]

    arXiv preprint arXiv:2408.12986 , year =

    Top Score on the Wrong Exam: On Benchmarking in Machine Learning for Vulnerability Detection , author =. arXiv preprint arXiv:2408.12986 , year =

    arXiv

  33. [41]

    IEEE Transactions on Software Engineering , year =

    Deep Learning Based Vulnerability Detection: Are We There Yet? , author =. IEEE Transactions on Software Engineering , year =

  34. [42]

    2024 , howpublished =

    2024

  35. [43]

    Jiang, Lingxiao and Misherghi, Ghassan and Su, Zhendong and Glondu, Stephane , booktitle =

  36. [44]

    Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods , author =. Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

  37. [45]

    Transactions of the Association for Computational Linguistics (TACL) , volume =

    Sparse, Dense, and Attentional Representations for Text Retrieval , author =. Transactions of the Association for Computational Linguistics (TACL) , volume =

  38. [46]

    Proceedings of the 2019 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE) , pages =

    When Deep Learning Met Code Search , author =. Proceedings of the 2019 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering (ESEC/FSE) , pages =

    2019

  39. [47]

    2012 , publisher =

    Serebryany, Konstantin and Bruening, Derek and Potapenko, Alexander and Vyukov, Dmitriy , booktitle =. 2012 , publisher =

    2012

  40. [48]

    2026 , howpublished =

    2026

  41. [49]

    2016 , publisher =

    Dolan-Gavitt, Brendan and Hulin, Patrick and Kirda, Engin and Leek, Tim and Mambretti, Andrea and Robertson, Wil and Ulrich, Frederick and Whelan, Ryan , booktitle =. 2016 , publisher =

    2016

  42. [50]

    2016 , publisher =

    Pewny, Jannik and Holz, Thorsten , booktitle =. 2016 , publisher =

    2016

  43. [51]

    Advances in Neural Information Processing Systems , volume =

    Training Language Models to Follow Instructions with Human Feedback , author =. Advances in Neural Information Processing Systems , volume =

  44. [52]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

  45. [53]

    mini-swe-agent: The 100 line AI agent that solves GitHub issues , year =

  46. [54]

    USENIX Security Symposium , year =

    On the Benefits of Cheating in Fuzzing: Reassessing Bug Injection for Fuzzer Evaluation , author =. USENIX Security Symposium , year =

  47. [55]

    arXiv preprint arXiv:2511.11019 , year =

    PatchEval: A Multi-Language Benchmark for Automated Vulnerability Repair , author =. arXiv preprint arXiv:2511.11019 , year =

  48. [56]

    2026 , month = jul, howpublished =

    2026

  49. [57]

    2026 , note =

    Luo, Xianzhen and Zhang, Jingyuan and Zhou, Shiqi and Huang, Rain and Xiao, Chuan and Zhu, Qingfu and Ma, Zhiyuan and Yue, Xing and Yue, Yang and Zeng, Wencong and Che, Wanxiang , booktitle =. 2026 , note =

    2026

  50. [58]

    , booktitle =

    Zelikman, Eric and Wu, Yuhuai and Mu, Jesse and Goodman, Noah D. , booktitle =

  51. [59]

    The Annals of Mathematical Statistics , volume =

    Table for Estimating the Goodness of Fit of Empirical Distributions , author =. The Annals of Mathematical Statistics , volume =

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.