Quick reintroduction, because a lot of you followed me for hakrawler and XSS payloads.
Since then: 3 kids, and a group of cyber companies. HackerContent runs social for cyber brands (some accounts you follow are ours!). Ironbark Cyber does offensive work. TRIAGERS™ handles bug bounty triage. I'm also a partner at Submersion AI, training small cyber models that hold their own against frontier ones.
These days most of my hacking runs through AI agents I built. Most of my consulting is helping businesses hand real work to AI without it blowing up.
That's what I'll post about.
Good to be back.
A new report from Acronis Threat Research Unit describes one operator running an APT-scale campaign across 11 countries using a 12-agent AI mesh wired into Kali, Ghidra, and Metasploit via MCP servers.
Per the report, commercial LLMs refused the offensive prompts, so the attacker switched to local abliterated models running on RTX 4090s.
They controlled the whole mesh through WeChat and LINE.
It wasn't clean fire-and-forget automation. The agents hallucinated, flooded the operator with noise, and the report says one agent's output indicated a military target fell outside its defined scope, just a guardrail doing its job, nothing more.
Guardrails on commercial models buy you five minutes at best. The real shift is how easily MCP lets one person wire local models straight into standard offensive tooling.
Full breakdown in the first comment.
You don't need to breach a Tier-1 bank's core infrastructure. You just need to pop one of their marketing vendors.
JPMorgan Chase has confirmed a breach after attackers hit external service providers.
Chase called this "marketing data," which sounds harmless in a press release but we know it's not.
I said it before: your perimeter is only as strong as your weakest vendor, and most companies have no idea what access their vendors actually have.
Pro tip: If you're vibe coding something, give it an established, well documented, opinionated framework to start with.
Something like Django or Laravel. Then ask it to follow the best practices, conventions and design patterns that are recommended by the framework documentation.
These frameworks already have excellent documentation and very clear conventions that will encourage your agent to write cleaner, more secure code by default, without needing a massive prompt.
Allow me to demonstrate why using CVSS to determine severity can be misleading. Each example gets more and more ridiculous. 👇
First vulnerability: Port 22 is open with SSH running.
CVSS breakdown:
Attack vector: Network
Attack complexity: Low (just log in)
Privileges required: High (requires a user account)
User interaction: None
Scope: Unchanged
Confidentiality: High
Integrity: High
Availability: High
Final severity: 7.2 (High)
I remember the first time I was hit with cyberattack.
It was September 9th, 2014. I woke up and an absolutely garbage U2 album had been downloaded straight to my iPhone 5.
Ask your AI agent "who do I know that would benefit from knowing each other" and then make the introduction.
It takes 30 seconds and could change someone's life!
@nmr031 If there was an RCE in a webapp that required the OS creds to work, it would be considered a vulnerability. What’s the difference?
Same impact, same CVSS ratings.
@nmr031 If there was an RCE in a webapp that required the OS creds to work, it would be considered a vulnerability. What’s the difference?
Same impact, same CVSS ratings.
@nmr031 If there was an RCE in a webapp that required the OS creds to work, it would be considered a vulnerability. What’s the difference?
Same impact, same CVSS ratings.