Skip to content
NewA calmer, faster ZeroHour

Fast-Tier Models Miss Held-Out Web Attack Families

Original title: Escalate or Close: How Fast-Tier Models Handle Held-Out Web Attack Families in Two-Tier Triage

arXiv cs.CRby Maciej Kisielewicz ()1 min read
AI summary · x-ai/grok-4.7Summaries by Vani AI

ModernBERT classifies 37-60% of held-out path traversal as benign within 5% budget; specialized models escalate 97.6% but may include public families in training.

In two-tier triage a fast model handles most requests while a slow tier examines only escalated ones. Researchers tested zero-shot decision models and supervised detectors against seven held-out web attack families by splicing payloads into real browsing traffic. ModernBERT gave held-out traversal a low probability far more often than when known, while Jev and Clef escalated nearly all within the same benign budget. Tests revealed request shape impacts results and benign twins bypass some models.

Unconfirmed: Training data of Jev and Clef undisclosed

Key facts
  • ModernBERT escalates 50-81% of held-out traversal within 5% benign budget
  • Jev and Clef escalate at least 97.6% of every held-out family
  • Jev misses 37% of benign twins with attack replaced by legitimate values
  • Specialized models' undisclosed training data may contain public families
  • Request shape controls needed for recent exploits
Analyze this story with Vani AIImpact, who is affected and what defenders should do, in a new chat

From the original article

1 min read at arxiv.org

In two-tier triage a fast model decides most web requests and escalates the rest to a slow tier, so an attack the fast tier closes is never examined.

arXiv cs.CR

By Maciej Kisielewicz · arxiv.org

Read the full article at arXiv cs.CR

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.