A software security harness mined fixed rules, patch clues and static signals across 4,245 C/C++ repositories. A small typed-decision model ranked bounded candidates instead of generating full exploit narratives; later reproducibility and human validation established which candidates were genuine. The researchers report 187 confirmed findings, with impact…
Inspect the ClaimsCase · DiggingBeagle record
A 4B typed decision model prioritized low-cost memory-safety research, while execution and human review supplied the actual proof
A software security harness mined fixed rules, patch clues and static signals across 4,245 C/C++ repositories. A small typed-decision model ranked bounded candidates instead of generating full exploit narratives; later reproducibility and human validation established which candidates were genuine.
Public disclosure: Nov 12, 2026
A defender has a static-analysis and executable-testing environment with many low-confidence security candidates.
At a glance
The model chooses investigation priority, while separately verifiable proof determines vulnerability and impact.
Read the implicationsThe AI mechanism is constrained decision-based triage, not giving an untrusted model authority to declare a flaw, assign a final severity or execute live exploits.
Limits and uncertaintyFull account
A software security harness mined fixed rules, patch clues and static signals across 4,245 C/C++ repositories. A small typed-decision model ranked bounded candidates instead of generating full exploit narratives; later reproducibility and human validation established which candidates were genuine.
The researchers report 187 confirmed findings, with impact categories and follow-up disclosure status. The model classifier itself had modest discrimination and weak pairwise vulnerable-versus-patched ranking; the pipeline required deterministic checks and substantial follow-through.
The 187 reports are not independently confirmed internet exploitation or 187 new CVEs with public patches. Reported throughput and cost omit parts of the original engineering effort; effectiveness depends on the candidate miner and validation budget.
Reported figures and denominator. 4,245 repositories examined, 187 researcher-reported findings; classifier AUC approximately 0.73, not a universal flaw detector.
The AI mechanism is constrained decision-based triage, not giving an untrusted model authority to declare a flaw, assign a final severity or execute live exploits.
Mechanism and trust boundary
- 01
Starting condition
A defender has a static-analysis and executable-testing environment with many low-confidence security candidates.
Boundary: precondition and attacker capability
- 02
Trust boundary crossed or evaluated
The researchers report 187 confirmed findings, with impact categories and follow-up disclosure status. The model classifier itself had modest discrimination and weak pairwise vulnerable-versus-patched ranking; the pipeline required deterministic checks and substantial follow-through.
Boundary: mechanism
- 03
Observed consequence
The model chooses investigation priority, while separately verifiable proof determines vulnerability and impact.
Boundary: observed effect and limitations
Claims & evidence
CLM-NOV15-DECISION4B-MECHA software security harness mined fixed rules, patch clues and static signals across 4,245 C/C++ repositories. A small typed-decision model ranked bounded candidates instead of generating full exploit narratives; later reproducibility and human validation established which candidates were genuine. The researchers report 187 confirmed findings, with impact categories and follow-up disclosure status. The model classifier itself had modest discrimination and weak pairwise vulnerable-versus-patched ranking; the pipeline required deterministic checks and substantial follow-through.supported
Basis: reported finding
- supportsSmall decision models used for memory safety vulnerability triage at scaleprimary disclosure
Decision-model paper, Method, Classification, campaign results and limitations
CLM-NOV15-DECISION4B-SCOPEThe 187 reports are not independently confirmed internet exploitation or 187 new CVEs with public patches. Reported throughput and cost omit parts of the original engineering effort; effectiveness depends on the candidate miner and validation budget. The AI mechanism is constrained decision-based triage, not giving an untrusted model authority to declare a flaw, assign a final severity or execute live exploits.supported
Basis: reported finding
- supportsSmall decision models used for memory safety vulnerability triage at scaleprimary disclosure
Primary disclosure: results, scope and limitations
Implications
The model chooses investigation priority, while separately verifiable proof determines vulnerability and impact.
Controls and mitigations
- Audit candidate selection recall and final executable test receipts rather than reporting ranked outputs as verified vulnerabilities.
- Retain human disclosure review and stage-wise compute/proof accounting so model cost claims do not omit testing.
Unknowns and contradictions
- The AI mechanism is constrained decision-based triage, not giving an untrusted model authority to declare a flaw, assign a final severity or execute live exploits.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 250