Evidence · DiggingBeagle record
CyberDual: AI Attackers versus AI Code Reviewers in Software Maintenance
Controlled malicious-patch corpus of 1,088 valid attempts and measurements against automated maintenance reviewers. The source differentiates detection by individual reviewer from the smaller set bypassing all reviewers.
- Source role
- primary disclosure
Evidence record
Controlled malicious-patch corpus of 1,088 valid attempts and measurements against automated maintenance reviewers. The source differentiates detection by individual reviewer from the smaller set bypassing all reviewers.
Claim-level citations (2)
- supportsMalicious maintenance-looking patches survived automated AI code review: The authors constructed 1,088 valid controlled attacks and found three agentic reviewers detected about 85% to 91% under their evaluation settings.
Abstract; Section 3 Results
- supportsMalicious maintenance-looking patches survived automated AI code review: The paper reports that roughly one in six maintenance-context attacks evaded particular reviewed configurations, distinct from code compromise in production repositories.
Section 3.1, 3.2; maintenance-window attack results