AI Assisted Decision Making of Security Review Needs for New Features
summary
Claim: natural language processing plus deep learning reliably flags which new features need security review. Implementation uses Apache Spark and Spark NLP, builds 300‑dimension word embeddings from project Jira/PRD text, compresses to a fixed 1,000‑word input and feeds a convolutional network. Early TF models hit 63–71% (ensemble 78%); the final model (clairo) reaches about 98% accuracy on ~700 test items and flags ~119 of ~800 features for review.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:09Security review bottlenecks in agile
- 3:35Security review decisions as automation
- 5:22Engineering vs spoken English
- 7:06Machine learning basics for classification
- 10:37First security review classifier attempt
- 15:35Term frequency and neural limits
- 16:37Unsupervised learning and conv nets
- 18:45Clo and Spark NLP implementation
- 24:46Model evaluation 98% accuracy
- 25:49Deploying model for real-world use
- 29:43Key takeaways and future of AI
- 31:06Q&A
related talks
23mAI assisted code reviews Quality and Security at Scale
48mAI Impact on Cybersecurity and Privacy
10mAI powered entomology: Lessons from millions of AI code reviews
42mAI Math Models Help Hunting Cybersecurity Threads
32mAI/Machine Learning: What is Actually Working in Cybersecurity?
28mSecuring AI by design
26mAI-powered Code Review
26mImproving Cybersecurity Posture with an AI Security Assistant