Step-by-Step Study Material

Step 1: The Problem — Measuring AI Cyber Capabilities

AI agents are increasingly able to find, exploit, and fix software vulnerabilities. Without a shared benchmark grounded in real systems, it is hard to quantify risk or progress.

Key idea: Evaluation should reflect evolving, real-world codebases—not only synthetic toy bugs.

Step 2: What BountyBench Provides

BountyBench includes:

• 25 systems with complex, real-world codebases
• 40 bug bounties covering 9 of the OWASP Top 10 web risks
• Monetary awards ranging from $10 to $30,485

These dollar amounts let researchers estimate the economic impact of agent successes and failures.

Step 3: The Vulnerability Lifecycle Tasks

To capture the full lifecycle, BountyBench defines three task types:

1. Detect — discover that a vulnerability exists
2. Exploit — demonstrate that the vulnerability can be abused
3. Patch — repair the issue while preserving system health

Difficulty can also be modulated by how much guidance is given, from near zero-day discovery to exploiting a known issue.

Step 4: What Early Agent Results Suggest

Evaluations of several AI agents show different offense/defense balances. Some coding agents are stronger at Patch than Exploit, while other custom agent setups are more balanced across offense and defense.

Takeaway: Useful cyber evaluation needs both offensive and defensive tasks, real systems, and metrics tied to practical impact.