I built an autonomous AI pentester. It had an 8-component architecture, used the latest LLMs, and was designed to find and exploit web application vulnerabilities without human guidance. Then I tested it head-to-head against ULTRA RED on standard vulnerable applications.
The results were humbling.
Testing environment: DVWA (Damn Vulnerable Web Application) and WebGoat, standard benchmarks for security testing tools.
Results comparison on vulnerability discovery:
Remote Code Execution (RCE): AI found 1, ULTRA RED found 27
SQL Injection (SQLi): AI found 1, ULTRA RED found 14
Cross-Site Scripting (XSS): AI found 3, ULTRA RED found 41
Beyond the numbers, the AI pentester:
Cost significantly more to run (LLM API costs add up fast during active scanning)
Took longer to complete each test cycle
Was less accurate — higher false positive and false negative rates
Failed to chain vulnerabilities the way a skilled attacker would
Lessons learned:
LLMs are planners, not evaluators. Language models are excellent at reasoning about what to try next, but poor at evaluating whether an attempt actually succeeded. Security testing requires tight feedback loops between action and validation that current LLMs handle poorly.
False negatives are deadly. Missing a real vulnerability is worse than finding a false positive. The AI's miss rate on RCE and SQLi was catastrophic from a security perspective.
AI lacks "hunger." Expert human hackers and well-engineered automated systems have a relentlessness that current AI agents struggle to replicate. They give up too easily when early attempts fail.
The future: AI as force multiplier, not replacement. AI can help security teams work faster and smarter — helping with threat analysis, generating remediation guidance, and augmenting human researchers. But replacing the core vulnerability detection and validation engine with a general-purpose LLM is not yet viable.
