AI on Kali Linux finds 12 flaws in 27 minutes on Metasploitable2

The experiment was set up in an isolated local network with Kali Linux as the attack platform and Metasploitable2 at 192.168.1.100 as the target. A three-tier Claude configuration (Opus, Sonnet, Haiku) was given only the target IP and then operated with no further human input until a review prompt appeared.

The AI proceeded through four phases: reconnaissance, scanning, exploitation, and post-exploitation. During reconnaissance it identified an old Linux kernel (2.6.x) and noted that such kernels often have local privilege escalation vulnerabilities. Scanning revealed over 30 open ports with outdated services, and the AI matched them to known CVEs with confidence scores, prioritizing a backdoored FTP service at 0.97 confidence. Exploitation yielded a root shell via that backdoor and a Tomcat manager login using default credentials, after which a Java shell was deployed. A human review was triggered for an SMB exploit because confidence was 0.63, below the 70% threshold; after approval, an SMB shell was obtained. Post-exploitation included reading /etc/passwd and /etc/shadow, and connecting to MySQL without a password.

The author notes that while the AI excelled at matching known CVEs to service versions and automating repetitive tasks like scanning and brute-forcing, it struggled with custom application logic, social engineering, and low-confidence decisions. The tool used, ALIEN PENTEST, is the author's own development and is presented with a three-day free trial.