AI Struggles with Complex Hacks in New Cybersecurity Benchmarks

AI can find some security flaws, but it's no match for skilled human hackers. That's the stark reality exposed by two new cybersecurity benchmarks: Cybench and CVE-Bench. These tests push AI to its limits in discovering and exploiting real-world software vulnerabilities, revealing both progress and significant shortcomings.

Cybench throws 40 professional-level capture-the-flag (CTF) challenges at AI models. These aren't toy problems; they're sourced from actual hacking competitions. The goal? Find a vulnerability, craft an exploit, and capture a hidden 'flag' string, mimicking a real attacker's process.

The benchmark's design is ruthlessly practical. Vulnerable systems run in sandboxed Docker containers. Success is binary: either the AI captures the flag or it fails. This mirrors the real world, where a partial exploit is often useless.

The results are sobering. Claude 3.5 Sonnet, the top performer, achieved only a 17.5% success rate in 'unguided' mode. More tellingly, every tested AI hit a hard wall. None could solve tasks that took human teams more than 11 minutes to crack. The gap between AI and skilled human hackers remains vast.

Cybench's 'subtask' mode breaks challenges into smaller steps, revealing where models falter in the exploit chain. This granular view exposes AI's struggle to connect disparate pieces of information, a critical skill in complex exploits.

CVE-Bench takes a different tack, focusing on 40 real, critical vulnerabilities from the National Vulnerability Database. These aren't contrived scenarios; they're serious flaws in open-source web applications that could lead to full system compromise.

The benchmark simulates both 'zero-day' scenarios (AI has no prior knowledge) and 'one-day' scenarios (vulnerability details are provided). Its grading system is clever: instead of looking for specific exploit code, it defines eight concrete outcomes representing a successful attack. This allows for creative solutions while ensuring the AI truly compromises the target.

What do these benchmarks tell us?

1. Testing AI security capabilities is possible in controlled, ethical ways. Containerization, clear success criteria, and real-world vulnerabilities create meaningful evaluations.

2. AI in offensive security is promising but limited. Models can solve some CTF challenges and exploit some known vulnerabilities, but they're far from consistently cracking difficult problems or finding novel exploits.

3. We now have a framework to measure progress. As AI models improve, these standardized tests will gauge their evolving capabilities in finding and exploiting security flaws.

The implications are significant. Defenders must understand AI's current limits when anticipating future threats. Researchers now have clear benchmarks to surpass as they push AI-assisted security tools forward.

The takeaway is clear: AI is becoming a factor in the security landscape, but it's not yet a replacement for skilled human attackers or defenders. These benchmarks give us a sobering reality check on AI's current capabilities, and a roadmap for its evolution in this critical domain.