OpenAI o3 finds CVE-2025–37899, a use-after-free in ksmbd

AI system

System: OpenAI o3
Model: o3.
Release: Exact release not established.
Application developer: OpenAI
Role: The model analyzed source code and identified a possible vulnerability for a researcher to verify.
Configuration: Not assessed
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: A code slice under roughly 10k-12k lines chosen by an expert. Repeated model runs with reports triaged by a human. A maintainer workflow that accepts and patches reports quickly.
Human oversight: Not assessed
Relation to AI: The model produced the finding unprompted during a benchmark. Heelan selected the code, validated the report and handled disclosure. He notes the model also produced false positives in other runs. The account is the independent deployer's own, corroborated by the CVE record and third-party coverage, not by OpenAI.
AI contribution: The model produced the finding unprompted during a benchmark. Heelan selected the code, validated the report and handled disclosure. He notes the model also produced false positives in other runs. The account is the independent deployer's own, corroborated by the CVE record and third-party coverage, not by OpenAI.

Outcome identifiers: Not assessed
Benefit: Sean Heelan, benchmarking o3 on ksmbd with no scaffolding, had the model surface an unknown use-after-free in the SMB logoff handler; the kernel CNA published CVE-2025–37899 on 20 May 2025.
Category: Reduced Risk
Status: Not assessed
Beneficiary groups: Society. Organizations.
Beneficiaries: Not assessed
Reach: Not assessed
Distribution and access: Not assessed
Duration: Not assessed
Expected?: Yes
Costs and harms: Not assessed

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: Not assessed
Comparison: Claude Sonnet 3.7 and Claude Sonnet 3.5 on the same 100-run benchmark (the known ksmbd Kerberos bug)
Results: Reference: Claude Sonnet 3.7 finds it 3 out of 100 runs and Claude Sonnet 3.5 does not find it in 100 runs. Result: o3 finds the kerberos authentication vulnerability in the benchmark in 8 of the 100 runs.
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Participants: Developer: OpenAI. Operator: Sean Heelan (independent researcher).
Jurisdictions: Developer: United States.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: Expert curation of context and validation of candidate bugs. Tolerance for false positives across runs.
Replication attempts: Not assessed
Later findings: 29 May 2025. Linuxiac reported that the upstream fix was merged quickly into every maintained kernel branch, and the CVE record lists five stable-branch commits.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: Society. Organizations.
Depth: Society: 2. Organizations: 1.
Reasons: Society: A remotely reachable kernel memory-safety bug was fixed across maintained branches before any reported exploitation. Exposure of ksmbd users is not quantified. Organizations: The kernel project received a report with a fix merged quickly, per Linuxiac.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: Independently reported

Kind: Not assessed
Related entries: Not assessed
Field: Cybersecurity / vulnerability research. Technology.
Purpose: General-purpose reasoning model used by the researcher to audit kernel source code for vulnerabilities.
Dates: Date basis unassessed: 2025 (precision unassessed). Updated: 21 October 2026.
Stage: Not assessed
Version: SHA-256: 9cb79554fbd3…