How I (my bots) went from developer to hacker in 2 hours

I'm a software developer and an architect. Used to do network and systems admin a long time ago. No security training, have never participated in a CTF, never wrote an exploit in my life.

Two hours before a CTF competition closed in late Feb 2026, I repurposed the AI coding agents I'd built for one of the challenges in the CTF and pointed them at the rest of the competition. I went from zero to fifth on the leaderboard before time ran out. When I showed the organisers how I did it, they kept the competition open to see how far I could push it.

Over the next two weeks I solved 57 challenges across web exploitation, reverse engineering, binary exploitation, cryptography, forensics, AI, and recon. I finished first individually. Three of those flags were captured only by the bots. Nobody else got them. In the final 24 hours the bots knocked out over 25 challenges. They were also handicapped: anything requiring physical presence was off the table. Only the digital/online flags was available to be captured.

The thing that made this work was not smarter models. It was the operating model around them. In a chat you ask a question and get an answer. In an agentic setup the agent holds state, picks up tools, remembers what failed, and keeps going until it gets somewhere. The agents wrote their own exploit scripts with pwntools, built GDB command files to pick apart binaries, automated SSH sessions into challenge networks with scripts, ran nmap scans, cracked encrypted archives with bkcrack, crafted SQL injection payloads, and used prompt injection to pull secrets out of LLM-gated apps. When a problem got too deep, specialist coding agents got pulled in. My job was picking targets and unsticking things when the environment broke.

That shift from prompt to goal is the part that matters for defenders. It changes attacker economics. Someone who can code but has never done security work can now run an offensive operation across multiple domains at speed, because the agentic loop handles the domain knowledge they do not have.

This talk covers how I set it up, how the agents operated, what broke, and what I learned. The focus is not on exploit mechanics but on how agent harnesses can discover and execute those techniques on their own, and why defenders need to understand this now.

A CTF is not the real world, but it is a clear signal of where things are heading. Since this work was done, Anthropic withheld its Mythos model from public release because its autonomous hacking capabilities were too dangerous, and OpenAI launched GPT-5.4-Cyber under tiered access for vetted defenders only. Six months sit between this submission and the conference in October. By then this threat may not be emerging anymore. This talk will incorporate the latest developments right up to the conference.

Content will be presented vendor-neutral consistent with CyberCon speaker guidelines.