One month into Claude 4.5, how are things hacking?

With the release of Claude 4.5 we saw tons of improvement upon its predecessor and even ran head-to-head benchmarks. Internally we actually had to update our test cases to make challenges harder in privilege escalation. In this article I want to go over what the state and capability of our hacking agent in late 2025.

“Dude, PTJ is finding some crazy sh#! on a web application.”

This was a message I received from a customer after we had made some significant engine changes to our hacking agent. There were 21 findings from this client penetration test including SSRF leading to read/write database access, LFI, apple private key disclosure and AWS credential disclosure. Note that there were zero false positives. General awareness of the LLMs has improved drastically in the last month and it’s clear Claude 4.5 is putting together logical chains previous models could not really put together. Below is an example from a run against a hard difficulty HackTheBox machine:

A previous model would probably have never discovered “SQL Server Linked Server Command Execution via xpcmdshell” in the first place, let alone use it to identify an internal linked server for privilege escalation. To be frank, we have never really seen something this dangerous before, and we deal with hacking agents all day. When a model becomes capable of persistent reasoning like we have here, human level penetration testing is in reach.

If you supply the agent with low privilege credentials on an Active Directory network, its ACL abuse capabilities are pretty spot on, and you have a good chance of watching it achieve Domain Admin while you sip your coffee. We designed the agent to analyze permissions by using the Json files created by running bloodhound-python, and it has been extremely successful. It is important to note that previous models like Claude 3.7 and 4 were unsuccessful in doing this task accurately and consistently. The future is definitely bright for autonomous hacking agents, and Vulnetic intends to lead the charge.