cai/docs/benchmarking/overview.md

174 lines
8.6 KiB
Markdown
Raw Permalink Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# Benchmarking Overview
CAIBench is a comprehensive meta-benchmark framework designed to rigorously evaluate cybersecurity AI agents across multiple domains. This framework enables standardized assessment of AI models and agents in both offensive and defensive security scenarios.
```
╔═══════════════════════════════════════════════════════════════════════════════╗
║ 🛡️ CAIBench Framework ⚔️ ║
║ Meta-benchmark Architecture ║
╚═══════════════════════════════════════════════════════════════════════════════╝
┌─────────────────────────────────┼────────────────────┐
│ │ │
🏛️ Categories 🚩 Difficulty 🐳 Infrastructure
│ │ │
┌─────────────────┼───────────────────┐ │ │
│ │ │ │ │ │ │
1⃣ 2⃣ 3⃣ 4⃣ 5⃣ │ │
Jeopardy A&D Cyber Knowledge Privacy │ Docker
CTF CTF Range Bench Bench │ Containers
│ │ │ │ │ │
┌──┴──┐ ┌──┴──┐ ┌──┴──┐ ┌──┴──┐ ┌──┴──┐ │
Base A&D Cyber SecEval CyberPII-Bench │
Cybench Ranges CTIBench │
RCTF2 CyberMetric │
AutoPenBench │
🚩───────🚩🚩───────🚩🚩🚩───────🚩🚩🚩🚩───────🚩🚩🚩🚩🚩
Beginner Novice Graduate Professional Elite
```
---
## 📊 Benchmark Results Overview
<table>
<tr>
<th style="text-align:center;"><b>Best Performance in Agent vs Agent A&D</b></th>
<th style="text-align:center;"><b>Model Performance in Jeopardy CTFs</b></th>
</tr>
<tr>
<td align="center"><img src="/assets/images/stackplot.png" alt="A&D Performance" width="100%" /></td>
<td align="center"><img src="/assets/images/base_1col.png" alt="Jeopardy CTF Performance" width="100%" /></td>
</tr>
<tr>
<th style="text-align:center;"><b>Model Performance in Privacy Benchmark</b></th>
<th style="text-align:center;"><b>Overall Model Performance</b></th>
</tr>
<tr>
<td align="center"><img src="/assets/images/cyberpii_benchmark.png" alt="Privacy Benchmark" width="100%" /></td>
<td align="center"><img src="/assets/images/caibench_spider.png" alt="Overall Performance" width="100%" /></td>
</tr>
</table>
**Key Insights from Benchmark Results:**
- 🥇 **alias1 dominates Attack & Defense CTFs** - Best offensive and defensive capabilities
- 🥇 **alias1 leads in Jeopardy-style CTFs** - Superior performance across all challenge types
- 🥇 **alias1 excels in privacy protection** - Highest F2 scores for PII handling
- 🥇 **alias1 shows balanced excellence** - Consistent top performance across all benchmark categories
---
## 🎯 What is CAIBench?
CAIBench is a **meta-benchmark** (benchmark of benchmarks) that:
- ✅ Evaluates AI agents across **offensive** and **defensive** security domains
- ✅ Uses **Docker containers** for reproducibility and isolation
- ✅ Provides **standardized metrics** for comparing AI models
- ✅ Covers **real-world scenarios** from CTFs, cyber ranges, and security operations
- ✅ Includes **privacy-aware** evaluation with PII handling benchmarks
---
## 📚 Research Foundation
CAIBench is backed by peer-reviewed research:
!!! tip "Core Research Papers"
📊 [**CAIBench: Cybersecurity AI Benchmark**](https://arxiv.org/pdf/2510.24317) (2025)
Modular meta-benchmark framework for evaluating LLM models and agents across offensive and defensive cybersecurity domains.
🎯 [**Evaluating Agentic Cybersecurity in Attack/Defense CTFs**](https://arxiv.org/pdf/2510.17521) (2025)
Real-world evaluation showing defensive agents achieved **54.3% patching success** versus **28.3% offensive initial access**.
**[View full research library →](https://aliasrobotics.com/research-security.php#papers)**
**[Browse benchmark source code →](https://github.com/aliasrobotics/cai/tree/main/benchmarks)**
---
## 🏆 Performance Highlights
### alias1 - Best-in-Class Performance
Based on CAIBench evaluations, **`alias1`** consistently outperforms all other models across cybersecurity benchmarks:
!!! success "alias1 Performance"
- 🥇 **#1 in Attack & Defense CTFs** - Superior offensive and defensive capabilities
- 🥇 **#1 in Jeopardy-style CTFs** - Best performance across web, pwn, crypto, forensics challenges
- 🥇 **#1 in Cyber Range scenarios** - Highest success rate in realistic environments
- 🥇 **Zero refusals** - Unrestricted responses for authorized security testing
**[See detailed benchmark results →](attack_defense.md)**
**[Learn more about alias1 →](../cai_pro.md)**
---
## 📊 Benchmark Categories
CAIBench evaluates AI agents across five categories:
### 1. Jeopardy-style CTFs
Independent challenges in cryptography, web exploitation, binary reversing, forensics, and more.
**[Learn more →](jeopardy_ctfs.md)**
### 2. Attack & Defense CTFs
Real-time competitive environments where agents must simultaneously attack opponents and defend their own systems.
**[Learn more →](attack_defense.md)**
### 3. Cyber Range Exercises
Realistic training environments with complex multi-system scenarios involving incident response and security operations.
**[Learn more →](cyber_ranges.md)**
### 4. Cybersecurity Knowledge
Question-answering benchmarks evaluating understanding of security concepts, threat intelligence, and vulnerability analysis.
**[Learn more →](knowledge_benchmarks.md)**
### 5. Privacy Benchmarks
Assessment of AI models' ability to handle sensitive information and properly manage Personally Identifiable Information (PII).
**[Learn more →](privacy_benchmarks.md)**
---
## 🚩 Difficulty Levels
Benchmarks are classified across five difficulty levels:
| Level | Persona | Target Audience |
|-------|---------|-----------------|
| 🚩 Very Easy | Beginner | High school students, cybersecurity beginners |
| 🚩🚩 Easy | Novice | Individuals familiar with basic security concepts |
| 🚩🚩🚩 Medium | Graduate Level | College students, security undergraduates/graduates |
| 🚩🚩🚩🚩 Hard | Professional | Working penetration testers, security professionals |
| 🚩🚩🚩🚩🚩 Very Hard | Elite | Advanced security researchers, elite participants |
---
## 🚀 Getting Started
Ready to run benchmarks? Check out:
- **[Running Benchmarks](running_benchmarks.md)** - Setup and usage instructions
- **[Attack & Defense Results](attack_defense.md)** - See alias1's superior performance
- **[GitHub Repository](https://github.com/aliasrobotics/cai/tree/main/benchmarks)** - Source code and examples
---
## 💡 Why Benchmarking Matters
Rigorous benchmarking is essential for:
- 📈 **Measuring Progress** - Track improvements in AI security capabilities over time
- 🔬 **Research Validation** - Provide scientific evidence for security AI effectiveness
- 🏆 **Model Comparison** - Enable objective comparison between AI models
- 🛡️ **Real-world Readiness** - Validate agents before deploying in production environments
- 🎓 **Educational Value** - Help researchers understand AI strengths and limitations
CAIBench provides the most comprehensive evaluation framework for cybersecurity AI, validated through peer-reviewed research and real-world CTF competitions.