Beyond Success Rate: Cost-Aware Evaluation of Offensive and Defensive Security Agents

Forum topic · 小凯 · 2026-07-18

Summary

This arXiv paper (2607.15263) by Paul Kassianik, Blaine Nelson, and Yaron Singer argues that security-agent benchmarks should measure cost-efficiency alongside task success. Standard evaluations report peak offensive capability under generous inference budgets (vulnerability discovery, exploit development, penetration testing, CTF completion), but in operational security every reasoning step, tool call, telemetry query, and enrichment request consumes budget. The authors evaluate language-model security agents on offensive Cybench challenges and defensive Splunk BOTS v1 investigation challenges, comparing models at fixed cost levels and decomposing performance by inference spend and tool spend. Results reveal distinct scaling regimes: offensive CTF performance improves with additional test-time compute, and scaled open-weight models can approach frontier proprietary systems while staying cost-competitive. Defensive SOC investigations scale differently—success depends more on disciplined tool use, telemetry navigation, and selective enrichment than on raw reasoning budget. The paper advocates cost-aware, SOC-native evaluation as a clearer picture of which models are practically useful today and where defensive agents still need improvement.

Paper Overview

Field: Machine Learning Authors: Paul Kassianik, Blaine Nelson, Yaron Singer Released: 2026-07-16 arXiv: 2607.15263

Abstract

Security-agent evaluations commonly measure peak offensive capability under generous inference budgets, emphasizing vulnerability discovery, exploit development, penetration testing, and CTF completion. Such measurements are useful but incomplete: in operational security, every reasoning step, tool call, telemetry query, and enrichment request consumes budget. The authors evaluate language-model security agents through this cost-success lens on offensive Cybench challenges and defensive Splunk BOTS v1 investigation challenges.

Instead of reporting only best-case success, they compare models at fixed cost levels and decompose performance by inference spend and tool spend.

Key Findings

Distinct scaling regimes for red- and blue-team tasks: offensive and defensive security workloads do not benefit from compute in the same way.
Offensive CTF performance improves with additional test-time compute; scaled open-weight models can approach frontier proprietary systems while remaining cost-competitive.
Defensive SOC investigations do not scale the same way: success depends more on disciplined tool use, telemetry navigation, and selective enrichment than on raw reasoning budget.

Recommendation

The authors argue that security-agent benchmarks should measure economic efficiency and operational fitness in addition to task success. Cost-aware, SOC-native evaluation provides a clearer picture of which models are practically useful today and where defensive agents still need improvement.