We ask whether small, open-weight decision models can support vulnerability discovery in open-source software at scale and what the surrounding system must provide. These models answer a typed question in one forward pass, returning scores for the permitted answers rather than text. Our harness selects widely packaged C and C++ projects and uses deterministic incomplete-fix miners and shape localizers to decide where to look. Intern-Decision-4B, a 4B-parameter open-weight decision model, then scores every candidate function. The scores rank the queues rather than accept or reject candidates, apart from a low reach-score floor that some queues kept. The search uses no frontier model. Alone, the model is weak: it separates vulnerable from benign functions with AUC 0.73, but a vulnerable function from its own patch with only 0.56. It is more useful for ordering candidates. On 75 public CVEs from projects without any of our findings, the fixed function reached its repository's top 25 in 21% of cases, 1.4 times as often as under random order of the same queues. Recall is bounded mainly by where the localizers look. On the campaign's own queues the model's order placed 7.5 times more confirmed findings in the top 10 than random order, an observed, selection-biased upper bound. On three GB10 systems, the harness searched more than 4,000 projects at 900-1,100 repositories per hour. The campaign confirmed 187 memory-safety findings; by worst-case CVSS v3.1, 12 are critical, 148 high, 23 medium and 4 low. The decision step runs at 4.4 decisions per second on one GB10 and 9.8 on a consumer RTX 4090 gaming PC, with matching scores. A small decision model supports vulnerability discovery through prioritization, while deterministic analysis and verification remain essential. All confirmed findings are being disclosed to maintainers under coordinated disclosure before publication.
Figures & tables
Configuration
AUC vs benign (native C/C++)
AUC vs patched (all 91 pairs)
Throughput on one GB10
4B + 27B two-tier (rank mean)
0.76
0.57
—
clef-27B
0.74
0.56
0.45 chunks/s
Intern-Decision-4B ⋅ memory-safety question
0.73
0.55
4.7 chunks/s
DiffusionGemma-26B-A4B
0.69
0.55
—
Table 1: Campaign deciders on the CVE function set. A dash marks a measurement not taken. DiffusionGemma-26B-A4B [ 135 ] was run as NVIDIA’s NVFP4 build [ 136 ] . The patched column covers all 91 pairs, including 20 in managed languages; on the 71 native pairs Intern scores 0.564 (Table 2 ).
Model
AUC vs benign
AUC vs patched
Recall at 0.5
Benign FPR at 0.5
Intern-Decision-4B
0.731
0.564
15%
6%
xLAM-2-1b-fc-r
0.773
0.590
73%
32%
xLAM-2-3b-fc-r
0.688
0.544
0%
0%
Phi-4-mini-instruct
0.666
0.617
79%
50%
Granite Guardian 3.1-2B
0.672
0.555
99%
84%
Granite Guardian 3.2-5B
0.561
0.504
34%
23%
Table 2: Small classifiers on the native C/C++ function set (71 vulnerable, 71 patched, 200 benign), one GB10 under vLLM. Intern’s 95% intervals are given in the text.
Model
Parameters
Readout
Rows/s
Intern-Decision-4B
4.5B
decision readout
4.7 *
xLAM-2-1b-fc-r
1.5B
Yes/No
49.1
xLAM-2-3b-fc-r
3.1B
Yes/No
25.3
Phi-4-mini-instruct
3.8B
Yes/No
23.1
Granite Guardian 3.1-2B
2.5B
guard template
18.8
Granite Guardian 3.2-5B
5.8B
guard template
10.0
Table 3: Configurations and throughput in the small-classifier comparison (one GB10, vLLM 0.30.1). * Intern answers two fields per row with longer prompts, so its rows/s is not comparable.
Candidate source
Random sample
Raw rate
Top k after ranking
Ranked rate
Fixed-buffer localizer
40 (first version)
2.5%
top 40 (revised version)
35–42.5% (16.7% without the validation repository)
Length-wrap localizer
30
6.7%
top 30
30–33%
Untrusted-size localizer
30
3.3%
top 25
24–36%
Table 4: Candidate yield in the localizer pilots, by configuration, at triage (descriptive). Raw rates are hit rates in a random sample of the localizer’s candidates; ranked rates are hit rates in the top k after the deterministic test, a reach cut at 0.27 and ranking by the shape question. Ranges count leads only and leads plus near-leads; they are not confidence intervals.
Stage (n = 75 CVEs)
Count
Share
Changed function in the chunked corpus
72
96%
Flagged by a localizer or miner (before caps)
29
39%
Scored (survived the per-repository caps)
18
24%
In the repository’s top 10 / 25 / 40
9 / 16 / 17
12% / 21% / 23%
Expected at random, same queues (top 10 / 25 / 40)
7.0 / 11.1 / 13.4
Table 5: Recall on 75 known CVEs. 95% intervals for recall at 10, 25 and 40: 5–20%, 12–31% and 13–32%.
Opengrep v1.31.0, C rules
Harness + 4B decision model
CPU wall time (same 16-vCPU host)
27–31 s at 16 jobs
≈ 268 s, single process
Decision step, one GB10, vLLM
—
278 decisions in 63 s (4.4/s)
Decision step, consumer RTX 4090, vLLM
—
278 decisions in 25–29 s (9.8/s)
Decision step, laptop integrated GPU (floor)
—
278 decisions in ≈ 10 min
End to end
≈ 0.5 min
≈ 5.5 min (GB10); ≈ 5 min (RTX 4090)
Output
554 warnings in 227 files
139 ranked rows; 5 where both questions agree
Table 6: One-repository cost comparison on OpenSSL. CPU stages ran on the same 16-vCPU x86 host. The harness’s 139 ranked rows cover 121 distinct functions. Neither output was triaged, so this is not a quality comparison.
Figure 2: Worst-case CVSS v3.1 ratings of the 187 confirmed findings, by severity class (score range in parentheses), from Critical to Low. Each bar is labelled with its count; the four classes total 187 findings. 160 of the 187 (86%) are rated critical or high under the worst-case rubric of Section 5.5 .
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
CVE
CWE
Repository
Fix commit
CVE-2021-33796
416
ccxvii/mujs
7ef066a3bb95
CVE-2021-45985
787
lua/lua
cf613cdc6fa3
CVE-2022-28071
416
radareorg/radare2
65448811e5b9
CVE-2022-48468
190
protobuf-c/protobuf-c
ec3d900001a1
CVE-2023-0341
787
editorconfig/editorconfig-core-c
41281ea82fbf
CVE-2023-2804
787
libjpeg-turbo/libjpeg-turbo
9f756bc67a84
Appendix
Table 7: The 75 public CVEs of the recall evaluation, with the fix commit used (the repository is checked out at its parent). Repositories are as referenced by NVD; one reference points to a GitHub mirror of the project.